diff --git a/README.md b/README.md index bb24b0f73..8ceec49fa 100644 --- a/README.md +++ b/README.md @@ -2,13 +2,13 @@ 机器人技术栈知识库 / Robotics research and engineering wiki. - + [![GitHub Pages](https://img.shields.io/badge/GitHub%20Pages-Live-brightgreen?logo=github)](https://imchong.github.io/Robotics_Notebooks/) [![Deploy GitHub Pages](https://github.com/ImChong/Robotics_Notebooks/actions/workflows/pages.yml/badge.svg)](https://github.com/ImChong/Robotics_Notebooks/actions/workflows/pages.yml) [![Wiki Lint](https://github.com/ImChong/Robotics_Notebooks/actions/workflows/lint.yml/badge.svg)](https://github.com/ImChong/Robotics_Notebooks/actions/workflows/lint.yml) [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](./LICENSE) -[![Knowledge Graph](https://img.shields.io/badge/知识图谱-1515节点_10187边-blue?logo=d3.js)](https://imchong.github.io/Robotics_Notebooks/graph.html) +[![Knowledge Graph](https://img.shields.io/badge/知识图谱-1513节点_10183边-blue?logo=d3.js)](https://imchong.github.io/Robotics_Notebooks/graph.html) [![Sources Coverage](https://img.shields.io/badge/sources覆盖率-98%25-green)](docs/checklists/tech-stack-next-phase-checklist-v26.md) diff --git a/docs/exports/graph-stats.json b/docs/exports/graph-stats.json index 9c531a59b..eff03b10c 100644 --- a/docs/exports/graph-stats.json +++ b/docs/exports/graph-stats.json @@ -1 +1 @@ -{"generated_at":"2026-06-29","node_count":1515,"edge_count":10187,"community_count":15,"top_hubs":[{"id":"wiki/overview/humanoid-paper-notebooks-index.md","label":"Humanoid Paper Notebooks 知识库索引","degree":396},{"id":"wiki/tasks/loco-manipulation.md","label":"Loco-Manipulation (移动操作)","degree":308},{"id":"wiki/concepts/sim2real.md","label":"Sim2Real","degree":238},{"id":"wiki/overview/paper-notebook-category-04-loco-manipulation-and-wbc.md","label":"Paper Notebooks · Loco-Manipulation and WBC","degree":234},{"id":"wiki/tasks/locomotion.md","label":"Locomotion","degree":216},{"id":"wiki/overview/humanoid-loco-manip-161-papers-technology-map.md","label":"人形 Loco-Manip 161 篇技术地图","degree":190},{"id":"wiki/methods/reinforcement-learning.md","label":"Reinforcement Learning (RL, 强化学习)","degree":185},{"id":"wiki/overview/humanoid-rl-motion-control-body-system-stack.md","label":"人形机器人 RL 运动控制：身体系统栈视角","degree":182},{"id":"wiki/methods/vla.md","label":"VLA（Vision-Language-Action）","degree":176},{"id":"wiki/tasks/manipulation.md","label":"Manipulation","degree":153}],"orphan_nodes":[],"type_distribution":{"entity":1036,"method":124,"overview":116,"concept":92,"query":57,"formalization":39,"comparison":37,"task":12,"reference":1,"roadmap_page":1},"community_distribution":{"其他（Other） 社区":266,"人形论文深读笔记（Humanoid Paper Notebooks） 社区":249,"移动操作（Loco-Manipulation） 社区":243,"视觉-语言-动作（VLA） 社区":129,"全身控制（Whole-Body Control, WBC） 社区":101,"操作（Manipulation） 社区":101,"模仿学习（Imitation Learning, IL） 社区":70,"人形机器人（Humanoid Robot） 社区":67,"物理引擎（MuJoCo） 社区":66,"人形 RL 运动控制（Humanoid RL Locomotion） 社区":64,"运动控制（Locomotion） 社区":44,"行为基础模型技术地图（BFM） 社区":44,"强化学习（Reinforcement Learning, RL） 社区":28,"宇树 G1 人形机器人（Unitree G1） 社区":26,"仿真到现实（Sim2Real） 社区":17},"community_quality":{"singleton_communities":[],"largest_community_ratio":0.164,"community_quality_warning":false},"latest_wiki_nodes":[{"path":"wiki/entities/paper-chord-contact-wrench-dexterous-manipulation.md","detail_id":"entity-paper-chord-contact-wrench-dexterous-manipulation","label":"CHORD（Contact Wrench Guidance for Dexterous Manipulation）","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/entities/paper-scenebot.md","detail_id":"entity-paper-scenebot","label":"SceneBot（Contact-Prompted Whole-Body Tracking with Scene-Interaction）","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/entities/paper-hapmorph-pneumatic-haptic-render.md","detail_id":"entity-paper-hapmorph-pneumatic-haptic-render","label":"HapMorph：多维气动触觉属性渲染框架","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/overview/topic-physics-fidelity.md","detail_id":"wiki-overview-topic-physics-fidelity","label":"仿真物理保真度（专题汇总）","type":"overview","recency":"2026-06-28","source":"log.md"},{"path":"wiki/queries/simulation-physics-fidelity.md","detail_id":"wiki-queries-simulation-physics-fidelity","label":"仿真物理保真度链路选型指南","type":"query","recency":"2026-06-28","source":"log.md"},{"path":"wiki/concepts/physics-fidelity-sim2real-gap.md","detail_id":"wiki-concepts-physics-fidelity-sim2real-gap","label":"Physics Fidelity ↔ Sim2Real Gap（物理保真度与仿真到现实差距）","type":"concept","recency":"2026-06-28","source":"log.md"},{"path":"wiki/entities/rek.md","detail_id":"entity-rek","label":"REK（Robot Embodied Kombat · 人形格斗联赛）","type":"entity","recency":"2026-06-28","source":"log.md"},{"path":"wiki/entities/gymnasium.md","detail_id":"entity-gymnasium","label":"Gymnasium（RL 环境 API 标准）","type":"entity","recency":"2026-06-28","source":"log.md"},{"path":"wiki/entities/paper-flap-fov-active-perception-3d-navigation.md","detail_id":"entity-paper-flap-fov-active-perception-3d-navigation","label":"FLAP（FOV 约束主动感知 · 无先验地图 3D 导航）","type":"entity","recency":"2026-06-28","source":"log.md"},{"path":"wiki/methods/newtons-method.md","detail_id":"wiki-methods-newtons-method","label":"Newton's Method（牛顿法）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/comparisons/second-order-optimizers.md","detail_id":"wiki-comparisons-second-order-optimizers","label":"Second-Order Optimizers：选型对比","type":"comparison","recency":"2026-06-27","source":"log.md"},{"path":"wiki/methods/sgd.md","detail_id":"wiki-methods-sgd","label":"SGD（Stochastic Gradient Descent）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/comparisons/deep-learning-optimizers.md","detail_id":"wiki-comparisons-deep-learning-optimizers","label":"Deep Learning Optimizers：选型对比","type":"comparison","recency":"2026-06-27","source":"log.md"},{"path":"wiki/entities/kinetiq-ascend.md","detail_id":"entity-kinetiq-ascend","label":"KinetIQ Ascend（Humanoid · 真机 VLA 强化学习后训练）","type":"entity","recency":"2026-06-27","source":"log.md"},{"path":"wiki/methods/vla.md","detail_id":"wiki-methods-vla","label":"VLA（Vision-Language-Action）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/methods/behavior-cloning.md","detail_id":"wiki-methods-behavior-cloning","label":"Behavior Cloning（行为克隆）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/tasks/manipulation.md","detail_id":"wiki-tasks-manipulation","label":"Manipulation","type":"task","recency":"2026-06-27","source":"log.md"},{"path":"wiki/concepts/3d-spatial-vqa.md","detail_id":"wiki-concepts-3d-spatial-vqa","label":"3D 空间 VQA（3D Spatial Visual Question Answering）","type":"concept","recency":"2026-06-26","source":"log.md"},{"path":"wiki/concepts/ai-auto-research.md","detail_id":"wiki-concepts-ai-auto-research","label":"AI Auto-Research（学术研究自动化）","type":"concept","recency":"2026-06-26","source":"log.md"},{"path":"wiki/concepts/armature-modeling.md","detail_id":"wiki-concepts-armature-modeling","label":"Armature Modeling（电枢惯量建模）","type":"concept","recency":"2026-06-26","source":"log.md"}],"latest_wiki_node":{"path":"wiki/entities/paper-chord-contact-wrench-dexterous-manipulation.md","detail_id":"entity-paper-chord-contact-wrench-dexterous-manipulation","label":"CHORD（Contact Wrench Guidance for Dexterous Manipulation）","type":"entity","recency":"2026-06-29","source":"log.md"}} \ No newline at end of file +{"generated_at":"2026-06-29","node_count":1513,"edge_count":10183,"community_count":15,"top_hubs":[{"id":"wiki/overview/humanoid-paper-notebooks-index.md","label":"Humanoid Paper Notebooks 知识库索引","degree":396},{"id":"wiki/tasks/loco-manipulation.md","label":"Loco-Manipulation (移动操作)","degree":307},{"id":"wiki/concepts/sim2real.md","label":"Sim2Real","degree":238},{"id":"wiki/overview/paper-notebook-category-04-loco-manipulation-and-wbc.md","label":"Paper Notebooks · Loco-Manipulation and WBC","degree":234},{"id":"wiki/tasks/locomotion.md","label":"Locomotion","degree":216},{"id":"wiki/overview/humanoid-loco-manip-161-papers-technology-map.md","label":"人形 Loco-Manip 161 篇技术地图","degree":189},{"id":"wiki/methods/reinforcement-learning.md","label":"Reinforcement Learning (RL, 强化学习)","degree":185},{"id":"wiki/overview/humanoid-rl-motion-control-body-system-stack.md","label":"人形机器人 RL 运动控制：身体系统栈视角","degree":182},{"id":"wiki/methods/vla.md","label":"VLA（Vision-Language-Action）","degree":176},{"id":"wiki/tasks/manipulation.md","label":"Manipulation","degree":153}],"orphan_nodes":[],"type_distribution":{"entity":1034,"method":124,"overview":116,"concept":92,"query":57,"formalization":39,"comparison":37,"task":12,"reference":1,"roadmap_page":1},"community_distribution":{"其他（Other） 社区":266,"人形论文深读笔记（Humanoid Paper Notebooks） 社区":249,"移动操作（Loco-Manipulation） 社区":241,"视觉-语言-动作（VLA） 社区":129,"全身控制（Whole-Body Control, WBC） 社区":101,"操作（Manipulation） 社区":101,"模仿学习（Imitation Learning, IL） 社区":70,"人形机器人（Humanoid Robot） 社区":67,"物理引擎（MuJoCo） 社区":66,"人形 RL 运动控制（Humanoid RL Locomotion） 社区":64,"运动控制（Locomotion） 社区":44,"行为基础模型技术地图（BFM） 社区":44,"强化学习（Reinforcement Learning, RL） 社区":28,"宇树 G1 人形机器人（Unitree G1） 社区":26,"仿真到现实（Sim2Real） 社区":17},"community_quality":{"singleton_communities":[],"largest_community_ratio":0.165,"community_quality_warning":false},"latest_wiki_nodes":[{"path":"wiki/entities/paper-sonic.md","detail_id":"entity-paper-sonic","label":"SONIC","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/methods/sonic-motion-tracking.md","detail_id":"wiki-methods-sonic-motion-tracking","label":"SONIC（规模化运动跟踪人形控制）","type":"method","recency":"2026-06-29","source":"log.md"},{"path":"wiki/entities/paper-chord-contact-wrench-dexterous-manipulation.md","detail_id":"entity-paper-chord-contact-wrench-dexterous-manipulation","label":"CHORD（Contact Wrench Guidance for Dexterous Manipulation）","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/entities/paper-scenebot.md","detail_id":"entity-paper-scenebot","label":"SceneBot（Contact-Prompted Whole-Body Tracking with Scene-Interaction）","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/entities/paper-hapmorph-pneumatic-haptic-render.md","detail_id":"entity-paper-hapmorph-pneumatic-haptic-render","label":"HapMorph：多维气动触觉属性渲染框架","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/overview/topic-physics-fidelity.md","detail_id":"wiki-overview-topic-physics-fidelity","label":"仿真物理保真度（专题汇总）","type":"overview","recency":"2026-06-28","source":"log.md"},{"path":"wiki/queries/simulation-physics-fidelity.md","detail_id":"wiki-queries-simulation-physics-fidelity","label":"仿真物理保真度链路选型指南","type":"query","recency":"2026-06-28","source":"log.md"},{"path":"wiki/concepts/physics-fidelity-sim2real-gap.md","detail_id":"wiki-concepts-physics-fidelity-sim2real-gap","label":"Physics Fidelity ↔ Sim2Real Gap（物理保真度与仿真到现实差距）","type":"concept","recency":"2026-06-28","source":"log.md"},{"path":"wiki/entities/rek.md","detail_id":"entity-rek","label":"REK（Robot Embodied Kombat · 人形格斗联赛）","type":"entity","recency":"2026-06-28","source":"log.md"},{"path":"wiki/entities/gymnasium.md","detail_id":"entity-gymnasium","label":"Gymnasium（RL 环境 API 标准）","type":"entity","recency":"2026-06-28","source":"log.md"},{"path":"wiki/entities/paper-flap-fov-active-perception-3d-navigation.md","detail_id":"entity-paper-flap-fov-active-perception-3d-navigation","label":"FLAP（FOV 约束主动感知 · 无先验地图 3D 导航）","type":"entity","recency":"2026-06-28","source":"log.md"},{"path":"wiki/methods/newtons-method.md","detail_id":"wiki-methods-newtons-method","label":"Newton's Method（牛顿法）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/comparisons/second-order-optimizers.md","detail_id":"wiki-comparisons-second-order-optimizers","label":"Second-Order Optimizers：选型对比","type":"comparison","recency":"2026-06-27","source":"log.md"},{"path":"wiki/methods/sgd.md","detail_id":"wiki-methods-sgd","label":"SGD（Stochastic Gradient Descent）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/comparisons/deep-learning-optimizers.md","detail_id":"wiki-comparisons-deep-learning-optimizers","label":"Deep Learning Optimizers：选型对比","type":"comparison","recency":"2026-06-27","source":"log.md"},{"path":"wiki/entities/kinetiq-ascend.md","detail_id":"entity-kinetiq-ascend","label":"KinetIQ Ascend（Humanoid · 真机 VLA 强化学习后训练）","type":"entity","recency":"2026-06-27","source":"log.md"},{"path":"wiki/methods/vla.md","detail_id":"wiki-methods-vla","label":"VLA（Vision-Language-Action）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/methods/behavior-cloning.md","detail_id":"wiki-methods-behavior-cloning","label":"Behavior Cloning（行为克隆）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/tasks/manipulation.md","detail_id":"wiki-tasks-manipulation","label":"Manipulation","type":"task","recency":"2026-06-27","source":"log.md"},{"path":"wiki/concepts/3d-spatial-vqa.md","detail_id":"wiki-concepts-3d-spatial-vqa","label":"3D 空间 VQA（3D Spatial Visual Question Answering）","type":"concept","recency":"2026-06-26","source":"log.md"}],"latest_wiki_node":{"path":"wiki/entities/paper-sonic.md","detail_id":"entity-paper-sonic","label":"SONIC","type":"entity","recency":"2026-06-29","source":"log.md"}} \ No newline at end of file diff --git a/docs/exports/home-stats.json b/docs/exports/home-stats.json index cbd67566e..87a7ebda9 100644 --- a/docs/exports/home-stats.json +++ b/docs/exports/home-stats.json @@ -1,13 +1,29 @@ { "generated_at": "2026-06-29", - "node_count": 1515, - "edge_count": 10187, + "node_count": 1513, + "edge_count": 10183, "coverage": { - "covered": 1484, - "total": 1513, + "covered": 1482, + "total": 1511, "percent": 98 }, "latest_wiki_nodes": [ + { + "path": "wiki/entities/paper-sonic.md", + "detail_id": "entity-paper-sonic", + "label": "SONIC", + "type": "entity", + "recency": "2026-06-29", + "source": "log.md" + }, + { + "path": "wiki/methods/sonic-motion-tracking.md", + "detail_id": "wiki-methods-sonic-motion-tracking", + "label": "SONIC（规模化运动跟踪人形控制）", + "type": "method", + "recency": "2026-06-29", + "source": "log.md" + }, { "path": "wiki/entities/paper-chord-contact-wrench-dexterous-manipulation.md", "detail_id": "entity-paper-chord-contact-wrench-dexterous-manipulation", @@ -151,28 +167,12 @@ "type": "concept", "recency": "2026-06-26", "source": "log.md" - }, - { - "path": "wiki/concepts/ai-auto-research.md", - "detail_id": "wiki-concepts-ai-auto-research", - "label": "AI Auto-Research（学术研究自动化）", - "type": "concept", - "recency": "2026-06-26", - "source": "log.md" - }, - { - "path": "wiki/concepts/armature-modeling.md", - "detail_id": "wiki-concepts-armature-modeling", - "label": "Armature Modeling（电枢惯量建模）", - "type": "concept", - "recency": "2026-06-26", - "source": "log.md" } ], "latest_wiki_node": { - "path": "wiki/entities/paper-chord-contact-wrench-dexterous-manipulation.md", - "detail_id": "entity-paper-chord-contact-wrench-dexterous-manipulation", - "label": "CHORD（Contact Wrench Guidance for Dexterous Manipulation）", + "path": "wiki/entities/paper-sonic.md", + "detail_id": "entity-paper-sonic", + "label": "SONIC", "type": "entity", "recency": "2026-06-29", "source": "log.md" diff --git a/docs/index.html b/docs/index.html index 6bca91772..9bcd31c7f 100644 --- a/docs/index.html +++ b/docs/index.html @@ -42,9 +42,9 @@

系统学习人形机器人运动控制

把 路线、图谱、模块、论文 串成一套极简导航系统。

- 1515 Nodes · - 10187 Links · - 1484/1513 Sources + 1513 Nodes · + 10183 Links · + 1482/1511 Sources
diff --git a/exports/graph-stats.json b/exports/graph-stats.json index 9c531a59b..eff03b10c 100644 --- a/exports/graph-stats.json +++ b/exports/graph-stats.json @@ -1 +1 @@ -{"generated_at":"2026-06-29","node_count":1515,"edge_count":10187,"community_count":15,"top_hubs":[{"id":"wiki/overview/humanoid-paper-notebooks-index.md","label":"Humanoid Paper Notebooks 知识库索引","degree":396},{"id":"wiki/tasks/loco-manipulation.md","label":"Loco-Manipulation (移动操作)","degree":308},{"id":"wiki/concepts/sim2real.md","label":"Sim2Real","degree":238},{"id":"wiki/overview/paper-notebook-category-04-loco-manipulation-and-wbc.md","label":"Paper Notebooks · Loco-Manipulation and WBC","degree":234},{"id":"wiki/tasks/locomotion.md","label":"Locomotion","degree":216},{"id":"wiki/overview/humanoid-loco-manip-161-papers-technology-map.md","label":"人形 Loco-Manip 161 篇技术地图","degree":190},{"id":"wiki/methods/reinforcement-learning.md","label":"Reinforcement Learning (RL, 强化学习)","degree":185},{"id":"wiki/overview/humanoid-rl-motion-control-body-system-stack.md","label":"人形机器人 RL 运动控制：身体系统栈视角","degree":182},{"id":"wiki/methods/vla.md","label":"VLA（Vision-Language-Action）","degree":176},{"id":"wiki/tasks/manipulation.md","label":"Manipulation","degree":153}],"orphan_nodes":[],"type_distribution":{"entity":1036,"method":124,"overview":116,"concept":92,"query":57,"formalization":39,"comparison":37,"task":12,"reference":1,"roadmap_page":1},"community_distribution":{"其他（Other） 社区":266,"人形论文深读笔记（Humanoid Paper Notebooks） 社区":249,"移动操作（Loco-Manipulation） 社区":243,"视觉-语言-动作（VLA） 社区":129,"全身控制（Whole-Body Control, WBC） 社区":101,"操作（Manipulation） 社区":101,"模仿学习（Imitation Learning, IL） 社区":70,"人形机器人（Humanoid Robot） 社区":67,"物理引擎（MuJoCo） 社区":66,"人形 RL 运动控制（Humanoid RL Locomotion） 社区":64,"运动控制（Locomotion） 社区":44,"行为基础模型技术地图（BFM） 社区":44,"强化学习（Reinforcement Learning, RL） 社区":28,"宇树 G1 人形机器人（Unitree G1） 社区":26,"仿真到现实（Sim2Real） 社区":17},"community_quality":{"singleton_communities":[],"largest_community_ratio":0.164,"community_quality_warning":false},"latest_wiki_nodes":[{"path":"wiki/entities/paper-chord-contact-wrench-dexterous-manipulation.md","detail_id":"entity-paper-chord-contact-wrench-dexterous-manipulation","label":"CHORD（Contact Wrench Guidance for Dexterous Manipulation）","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/entities/paper-scenebot.md","detail_id":"entity-paper-scenebot","label":"SceneBot（Contact-Prompted Whole-Body Tracking with Scene-Interaction）","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/entities/paper-hapmorph-pneumatic-haptic-render.md","detail_id":"entity-paper-hapmorph-pneumatic-haptic-render","label":"HapMorph：多维气动触觉属性渲染框架","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/overview/topic-physics-fidelity.md","detail_id":"wiki-overview-topic-physics-fidelity","label":"仿真物理保真度（专题汇总）","type":"overview","recency":"2026-06-28","source":"log.md"},{"path":"wiki/queries/simulation-physics-fidelity.md","detail_id":"wiki-queries-simulation-physics-fidelity","label":"仿真物理保真度链路选型指南","type":"query","recency":"2026-06-28","source":"log.md"},{"path":"wiki/concepts/physics-fidelity-sim2real-gap.md","detail_id":"wiki-concepts-physics-fidelity-sim2real-gap","label":"Physics Fidelity ↔ Sim2Real Gap（物理保真度与仿真到现实差距）","type":"concept","recency":"2026-06-28","source":"log.md"},{"path":"wiki/entities/rek.md","detail_id":"entity-rek","label":"REK（Robot Embodied Kombat · 人形格斗联赛）","type":"entity","recency":"2026-06-28","source":"log.md"},{"path":"wiki/entities/gymnasium.md","detail_id":"entity-gymnasium","label":"Gymnasium（RL 环境 API 标准）","type":"entity","recency":"2026-06-28","source":"log.md"},{"path":"wiki/entities/paper-flap-fov-active-perception-3d-navigation.md","detail_id":"entity-paper-flap-fov-active-perception-3d-navigation","label":"FLAP（FOV 约束主动感知 · 无先验地图 3D 导航）","type":"entity","recency":"2026-06-28","source":"log.md"},{"path":"wiki/methods/newtons-method.md","detail_id":"wiki-methods-newtons-method","label":"Newton's Method（牛顿法）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/comparisons/second-order-optimizers.md","detail_id":"wiki-comparisons-second-order-optimizers","label":"Second-Order Optimizers：选型对比","type":"comparison","recency":"2026-06-27","source":"log.md"},{"path":"wiki/methods/sgd.md","detail_id":"wiki-methods-sgd","label":"SGD（Stochastic Gradient Descent）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/comparisons/deep-learning-optimizers.md","detail_id":"wiki-comparisons-deep-learning-optimizers","label":"Deep Learning Optimizers：选型对比","type":"comparison","recency":"2026-06-27","source":"log.md"},{"path":"wiki/entities/kinetiq-ascend.md","detail_id":"entity-kinetiq-ascend","label":"KinetIQ Ascend（Humanoid · 真机 VLA 强化学习后训练）","type":"entity","recency":"2026-06-27","source":"log.md"},{"path":"wiki/methods/vla.md","detail_id":"wiki-methods-vla","label":"VLA（Vision-Language-Action）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/methods/behavior-cloning.md","detail_id":"wiki-methods-behavior-cloning","label":"Behavior Cloning（行为克隆）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/tasks/manipulation.md","detail_id":"wiki-tasks-manipulation","label":"Manipulation","type":"task","recency":"2026-06-27","source":"log.md"},{"path":"wiki/concepts/3d-spatial-vqa.md","detail_id":"wiki-concepts-3d-spatial-vqa","label":"3D 空间 VQA（3D Spatial Visual Question Answering）","type":"concept","recency":"2026-06-26","source":"log.md"},{"path":"wiki/concepts/ai-auto-research.md","detail_id":"wiki-concepts-ai-auto-research","label":"AI Auto-Research（学术研究自动化）","type":"concept","recency":"2026-06-26","source":"log.md"},{"path":"wiki/concepts/armature-modeling.md","detail_id":"wiki-concepts-armature-modeling","label":"Armature Modeling（电枢惯量建模）","type":"concept","recency":"2026-06-26","source":"log.md"}],"latest_wiki_node":{"path":"wiki/entities/paper-chord-contact-wrench-dexterous-manipulation.md","detail_id":"entity-paper-chord-contact-wrench-dexterous-manipulation","label":"CHORD（Contact Wrench Guidance for Dexterous Manipulation）","type":"entity","recency":"2026-06-29","source":"log.md"}} \ No newline at end of file +{"generated_at":"2026-06-29","node_count":1513,"edge_count":10183,"community_count":15,"top_hubs":[{"id":"wiki/overview/humanoid-paper-notebooks-index.md","label":"Humanoid Paper Notebooks 知识库索引","degree":396},{"id":"wiki/tasks/loco-manipulation.md","label":"Loco-Manipulation (移动操作)","degree":307},{"id":"wiki/concepts/sim2real.md","label":"Sim2Real","degree":238},{"id":"wiki/overview/paper-notebook-category-04-loco-manipulation-and-wbc.md","label":"Paper Notebooks · Loco-Manipulation and WBC","degree":234},{"id":"wiki/tasks/locomotion.md","label":"Locomotion","degree":216},{"id":"wiki/overview/humanoid-loco-manip-161-papers-technology-map.md","label":"人形 Loco-Manip 161 篇技术地图","degree":189},{"id":"wiki/methods/reinforcement-learning.md","label":"Reinforcement Learning (RL, 强化学习)","degree":185},{"id":"wiki/overview/humanoid-rl-motion-control-body-system-stack.md","label":"人形机器人 RL 运动控制：身体系统栈视角","degree":182},{"id":"wiki/methods/vla.md","label":"VLA（Vision-Language-Action）","degree":176},{"id":"wiki/tasks/manipulation.md","label":"Manipulation","degree":153}],"orphan_nodes":[],"type_distribution":{"entity":1034,"method":124,"overview":116,"concept":92,"query":57,"formalization":39,"comparison":37,"task":12,"reference":1,"roadmap_page":1},"community_distribution":{"其他（Other） 社区":266,"人形论文深读笔记（Humanoid Paper Notebooks） 社区":249,"移动操作（Loco-Manipulation） 社区":241,"视觉-语言-动作（VLA） 社区":129,"全身控制（Whole-Body Control, WBC） 社区":101,"操作（Manipulation） 社区":101,"模仿学习（Imitation Learning, IL） 社区":70,"人形机器人（Humanoid Robot） 社区":67,"物理引擎（MuJoCo） 社区":66,"人形 RL 运动控制（Humanoid RL Locomotion） 社区":64,"运动控制（Locomotion） 社区":44,"行为基础模型技术地图（BFM） 社区":44,"强化学习（Reinforcement Learning, RL） 社区":28,"宇树 G1 人形机器人（Unitree G1） 社区":26,"仿真到现实（Sim2Real） 社区":17},"community_quality":{"singleton_communities":[],"largest_community_ratio":0.165,"community_quality_warning":false},"latest_wiki_nodes":[{"path":"wiki/entities/paper-sonic.md","detail_id":"entity-paper-sonic","label":"SONIC","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/methods/sonic-motion-tracking.md","detail_id":"wiki-methods-sonic-motion-tracking","label":"SONIC（规模化运动跟踪人形控制）","type":"method","recency":"2026-06-29","source":"log.md"},{"path":"wiki/entities/paper-chord-contact-wrench-dexterous-manipulation.md","detail_id":"entity-paper-chord-contact-wrench-dexterous-manipulation","label":"CHORD（Contact Wrench Guidance for Dexterous Manipulation）","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/entities/paper-scenebot.md","detail_id":"entity-paper-scenebot","label":"SceneBot（Contact-Prompted Whole-Body Tracking with Scene-Interaction）","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/entities/paper-hapmorph-pneumatic-haptic-render.md","detail_id":"entity-paper-hapmorph-pneumatic-haptic-render","label":"HapMorph：多维气动触觉属性渲染框架","type":"entity","recency":"2026-06-29","source":"log.md"},{"path":"wiki/overview/topic-physics-fidelity.md","detail_id":"wiki-overview-topic-physics-fidelity","label":"仿真物理保真度（专题汇总）","type":"overview","recency":"2026-06-28","source":"log.md"},{"path":"wiki/queries/simulation-physics-fidelity.md","detail_id":"wiki-queries-simulation-physics-fidelity","label":"仿真物理保真度链路选型指南","type":"query","recency":"2026-06-28","source":"log.md"},{"path":"wiki/concepts/physics-fidelity-sim2real-gap.md","detail_id":"wiki-concepts-physics-fidelity-sim2real-gap","label":"Physics Fidelity ↔ Sim2Real Gap（物理保真度与仿真到现实差距）","type":"concept","recency":"2026-06-28","source":"log.md"},{"path":"wiki/entities/rek.md","detail_id":"entity-rek","label":"REK（Robot Embodied Kombat · 人形格斗联赛）","type":"entity","recency":"2026-06-28","source":"log.md"},{"path":"wiki/entities/gymnasium.md","detail_id":"entity-gymnasium","label":"Gymnasium（RL 环境 API 标准）","type":"entity","recency":"2026-06-28","source":"log.md"},{"path":"wiki/entities/paper-flap-fov-active-perception-3d-navigation.md","detail_id":"entity-paper-flap-fov-active-perception-3d-navigation","label":"FLAP（FOV 约束主动感知 · 无先验地图 3D 导航）","type":"entity","recency":"2026-06-28","source":"log.md"},{"path":"wiki/methods/newtons-method.md","detail_id":"wiki-methods-newtons-method","label":"Newton's Method（牛顿法）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/comparisons/second-order-optimizers.md","detail_id":"wiki-comparisons-second-order-optimizers","label":"Second-Order Optimizers：选型对比","type":"comparison","recency":"2026-06-27","source":"log.md"},{"path":"wiki/methods/sgd.md","detail_id":"wiki-methods-sgd","label":"SGD（Stochastic Gradient Descent）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/comparisons/deep-learning-optimizers.md","detail_id":"wiki-comparisons-deep-learning-optimizers","label":"Deep Learning Optimizers：选型对比","type":"comparison","recency":"2026-06-27","source":"log.md"},{"path":"wiki/entities/kinetiq-ascend.md","detail_id":"entity-kinetiq-ascend","label":"KinetIQ Ascend（Humanoid · 真机 VLA 强化学习后训练）","type":"entity","recency":"2026-06-27","source":"log.md"},{"path":"wiki/methods/vla.md","detail_id":"wiki-methods-vla","label":"VLA（Vision-Language-Action）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/methods/behavior-cloning.md","detail_id":"wiki-methods-behavior-cloning","label":"Behavior Cloning（行为克隆）","type":"method","recency":"2026-06-27","source":"log.md"},{"path":"wiki/tasks/manipulation.md","detail_id":"wiki-tasks-manipulation","label":"Manipulation","type":"task","recency":"2026-06-27","source":"log.md"},{"path":"wiki/concepts/3d-spatial-vqa.md","detail_id":"wiki-concepts-3d-spatial-vqa","label":"3D 空间 VQA（3D Spatial Visual Question Answering）","type":"concept","recency":"2026-06-26","source":"log.md"}],"latest_wiki_node":{"path":"wiki/entities/paper-sonic.md","detail_id":"entity-paper-sonic","label":"SONIC","type":"entity","recency":"2026-06-29","source":"log.md"}} \ No newline at end of file diff --git a/exports/home-stats.json b/exports/home-stats.json index cbd67566e..87a7ebda9 100644 --- a/exports/home-stats.json +++ b/exports/home-stats.json @@ -1,13 +1,29 @@ { "generated_at": "2026-06-29", - "node_count": 1515, - "edge_count": 10187, + "node_count": 1513, + "edge_count": 10183, "coverage": { - "covered": 1484, - "total": 1513, + "covered": 1482, + "total": 1511, "percent": 98 }, "latest_wiki_nodes": [ + { + "path": "wiki/entities/paper-sonic.md", + "detail_id": "entity-paper-sonic", + "label": "SONIC", + "type": "entity", + "recency": "2026-06-29", + "source": "log.md" + }, + { + "path": "wiki/methods/sonic-motion-tracking.md", + "detail_id": "wiki-methods-sonic-motion-tracking", + "label": "SONIC（规模化运动跟踪人形控制）", + "type": "method", + "recency": "2026-06-29", + "source": "log.md" + }, { "path": "wiki/entities/paper-chord-contact-wrench-dexterous-manipulation.md", "detail_id": "entity-paper-chord-contact-wrench-dexterous-manipulation", @@ -151,28 +167,12 @@ "type": "concept", "recency": "2026-06-26", "source": "log.md" - }, - { - "path": "wiki/concepts/ai-auto-research.md", - "detail_id": "wiki-concepts-ai-auto-research", - "label": "AI Auto-Research（学术研究自动化）", - "type": "concept", - "recency": "2026-06-26", - "source": "log.md" - }, - { - "path": "wiki/concepts/armature-modeling.md", - "detail_id": "wiki-concepts-armature-modeling", - "label": "Armature Modeling（电枢惯量建模）", - "type": "concept", - "recency": "2026-06-26", - "source": "log.md" } ], "latest_wiki_node": { - "path": "wiki/entities/paper-chord-contact-wrench-dexterous-manipulation.md", - "detail_id": "entity-paper-chord-contact-wrench-dexterous-manipulation", - "label": "CHORD（Contact Wrench Guidance for Dexterous Manipulation）", + "path": "wiki/entities/paper-sonic.md", + "detail_id": "entity-paper-sonic", + "label": "SONIC", "type": "entity", "recency": "2026-06-29", "source": "log.md" diff --git a/exports/lint-report.md b/exports/lint-report.md index 1c9d16d62..7c626d03d 100644 --- a/exports/lint-report.md +++ b/exports/lint-report.md @@ -107,4 +107,4 @@ ### 💡 动力学/仿真/物理概念页缺回链「仿真物理保真度」专题枢纽（信息型，不阻塞 CI）（0 个） - 无 -📊 Sources 覆盖率：1484/1513 (98%) wiki/entity 页有 ingest 来源 +📊 Sources 覆盖率：1482/1511 (98%) wiki/entity 页有 ingest 来源 diff --git a/log.md b/log.md index 7ab86dc1d..38432ecac 100644 --- a/log.md +++ b/log.md @@ -1,5 +1,11 @@ > 核心规范：所有日常动作（ingest / query / lint / structural）必须追加记录到此文件。 +## [2026-06-29] structural | wiki/entities/paper-sonic.md — 合并 Loco-Manip 161 重复 SONIC stub（#019/#103）至 canonical 实体 + 方法页 + +- 删除：[`paper-loco-manip-161-019-sonic.md`](wiki/entities/paper-loco-manip-161-019-sonic.md)、[`paper-loco-manip-161-103-sonic.md`](wiki/entities/paper-loco-manip-161-103-sonic.md) +- 保留 canonical：[`paper-sonic.md`](wiki/entities/paper-sonic.md) + [`sonic-motion-tracking.md`](wiki/methods/sonic-motion-tracking.md) +- 交叉更新：Loco-Manip 161 category 01/04 表、[`humanoid_loco_manip_161_catalog.md`](sources/papers/humanoid_loco_manip_161_catalog.md)、对应 source 映射；[`bootstrap_loco_manip_161_entities.py`](scripts/bootstrap_loco_manip_161_entities.py) 增加 `CANONICAL_ENTITY_BY_NUM` 防再生 + ## [2026-06-29] ingest | sources/papers/chord_nvidia_video_to_data_2026.md — CHORD 接触力旋量引导灵巧操作；wiki/entities/paper-chord-contact-wrench-dexterous-manipulation.md；交叉 contact-rich-manipulation / manipulation / SPIDER / dexterous-data-pipeline / Isaac Lab ## [2026-06-29] ingest | sources/papers/scenebot_arxiv_2606_27581.md — 接入 SceneBot contact-prompted 全身场景交互跟踪；沉淀 wiki/entities/paper-scenebot.md；交叉更新 SONIC、运动跟踪选型、loco-manipulation、OmniRetarget diff --git a/scripts/bootstrap_loco_manip_161_entities.py b/scripts/bootstrap_loco_manip_161_entities.py index 07e8b7176..b16b664f6 100644 --- a/scripts/bootstrap_loco_manip_161_entities.py +++ b/scripts/bootstrap_loco_manip_161_entities.py @@ -39,6 +39,12 @@ def _resolve_raw_md() -> Path: # Prior catalog cross-links (wiki path without .md) for related pages PRIOR_WIKI: dict[int, str] = {} +# Loco-Manip 161 槽位 → 已有 canonical 实体（跳过独立 stub 生成） +CANONICAL_ENTITY_BY_NUM: dict[int, str] = { + 19: "paper-sonic", + 103: "paper-sonic", +} + def slugify(text: str, num: int) -> str: s = re.sub(r"[^a-zA-Z0-9]+", "-", text).strip("-").lower() @@ -394,6 +400,11 @@ def main() -> None: assert len(papers) == 161, len(papers) for p in papers: + canonical = CANONICAL_ENTITY_BY_NUM.get(p["num"]) + if canonical: + p["entity"] = canonical + write_source(p, canonical) + continue prior = prior_map.get(p["num"]) en = write_entity(p, prior) write_source(p, en) diff --git a/sources/papers/humanoid_loco_manip_161_catalog.md b/sources/papers/humanoid_loco_manip_161_catalog.md index 8af2ced11..2ff477568 100644 --- a/sources/papers/humanoid_loco_manip_161_catalog.md +++ b/sources/papers/humanoid_loco_manip_161_catalog.md @@ -32,7 +32,7 @@ | 016 | OmniH2O | [paper-loco-manip-161-016-omnih2o](../../wiki/entities/paper-loco-manip-161-016-omnih2o.md) | | 017 | OmniRetarget | [paper-loco-manip-161-017-omniretarget](../../wiki/entities/paper-loco-manip-161-017-omniretarget.md) | | 018 | Retargeting | [paper-loco-manip-161-018-retargeting](../../wiki/entities/paper-loco-manip-161-018-retargeting.md) | -| 019 | SONIC | [paper-loco-manip-161-019-sonic](../../wiki/entities/paper-loco-manip-161-019-sonic.md) | +| 019 | SONIC | [paper-sonic](../../wiki/entities/paper-sonic.md) | | 020 | TWIST2 | [paper-loco-manip-161-020-twist2](../../wiki/entities/paper-loco-manip-161-020-twist2.md) | | 021 | TWIST | [paper-loco-manip-161-021-twist](../../wiki/entities/paper-loco-manip-161-021-twist.md) | | 022 | TextOp | [paper-loco-manip-161-022-textop](../../wiki/entities/paper-loco-manip-161-022-textop.md) | @@ -137,7 +137,7 @@ | 100 | MotionWAM | [paper-loco-manip-161-100-motionwam](../../wiki/entities/paper-loco-manip-161-100-motionwam.md) | | 101 | OMG | [paper-loco-manip-161-101-omg](../../wiki/entities/paper-loco-manip-161-101-omg.md) | | 102 | Robot Motion Diffusion Model | [paper-loco-manip-161-102-robot-motion-diffusion-model](../../wiki/entities/paper-loco-manip-161-102-robot-motion-diffusion-model.md) | -| 103 | SONIC | [paper-loco-manip-161-103-sonic](../../wiki/entities/paper-loco-manip-161-103-sonic.md) | +| 103 | SONIC | [paper-sonic](../../wiki/entities/paper-sonic.md) | | 104 | SafeFlow | [paper-loco-manip-161-104-safeflow](../../wiki/entities/paper-loco-manip-161-104-safeflow.md) | | 105 | TextOp | [paper-loco-manip-161-105-textop](../../wiki/entities/paper-loco-manip-161-105-textop.md) | | 106 | 从语言到运动 | [paper-loco-manip-161-106-n106](../../wiki/entities/paper-loco-manip-161-106-n106.md) | diff --git a/sources/papers/loco_manip_161_survey_019_sonic.md b/sources/papers/loco_manip_161_survey_019_sonic.md index bac235c3f..c4abc3a6f 100644 --- a/sources/papers/loco_manip_161_survey_019_sonic.md +++ b/sources/papers/loco_manip_161_survey_019_sonic.md @@ -18,7 +18,7 @@ ## 对 wiki 的映射 -- [paper-loco-manip-161-019-sonic](../../wiki/entities/paper-loco-manip-161-019-sonic.md) +- [paper-sonic](../../wiki/entities/paper-sonic.md) - [loco-manip-161-category-01-motion-base-wbt](../../wiki/overview/loco-manip-161-category-01-motion-base-wbt.md) ## 参考来源（原始） diff --git a/sources/papers/loco_manip_161_survey_103_sonic.md b/sources/papers/loco_manip_161_survey_103_sonic.md index 4963e009f..42e467187 100644 --- a/sources/papers/loco_manip_161_survey_103_sonic.md +++ b/sources/papers/loco_manip_161_survey_103_sonic.md @@ -18,7 +18,7 @@ ## 对 wiki 的映射 -- [paper-loco-manip-161-103-sonic](../../wiki/entities/paper-loco-manip-161-103-sonic.md) +- [paper-sonic](../../wiki/entities/paper-sonic.md) - [loco-manip-161-category-04-generative-language-trajectory](../../wiki/overview/loco-manip-161-category-04-generative-language-trajectory.md) ## 参考来源（原始） diff --git a/wiki/entities/paper-loco-manip-161-001-agility-meets-stability.md b/wiki/entities/paper-loco-manip-161-001-agility-meets-stability.md index ca8266558..ac2a05936 100644 --- a/wiki/entities/paper-loco-manip-161-001-agility-meets-stability.md +++ b/wiki/entities/paper-loco-manip-161-001-agility-meets-stability.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Agility Meets Stability 把本体状态与关节序列、人类视频/动捕轨迹、仿真交互数据转成可跟踪的身体目标，并通过异构动捕与合成平衡数据、混合奖励与动态采样、分层技能/专家策略训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把敏捷动捕和合成平衡样本放进同一策略，但用混合奖励分别约束动作跟踪与极端平衡。" related: diff --git a/wiki/entities/paper-loco-manip-161-002-any2any.md b/wiki/entities/paper-loco-manip-161-002-any2any.md index 351b10e8d..0e171a742 100644 --- a/wiki/entities/paper-loco-manip-161-002-any2any.md +++ b/wiki/entities/paper-loco-manip-161-002-any2any.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Any2Any 把视觉、状态和动作数据转成可跟踪的身体目标，并通过源/目标机器人运动学对齐、PEFT 动力学适配、全身控制器/WBC/MPC训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是先对齐源/目标机器人的状态和动作空间，再只微调动力学敏感模块，尽量保留原策略的运动先验。" related: diff --git a/wiki/entities/paper-loco-manip-161-003-bfm-zero.md b/wiki/entities/paper-loco-manip-161-003-bfm-zero.md index 57d882614..9754ba828 100644 --- a/wiki/entities/paper-loco-manip-161-003-bfm-zero.md +++ b/wiki/entities/paper-loco-manip-161-003-bfm-zero.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "BFM-Zero 把本体状态与关节序列、仿真交互数据、接触力/触觉信号转成可跟踪的身体目标，并通过PPO/RL 策略训练、分层技能/专家策略、闭环纠错/人类干预训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-004-beyondmimic.md b/wiki/entities/paper-loco-manip-161-004-beyondmimic.md index 2eadca8ef..165a481dd 100644 --- a/wiki/entities/paper-loco-manip-161-004-beyondmimic.md +++ b/wiki/entities/paper-loco-manip-161-004-beyondmimic.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "BeyondMimic 把人类视频/动捕轨迹转成可跟踪的身体目标，并通过ACT/行为克隆模仿学习、扩散策略/流匹配、分层技能/专家策略训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-005-chip.md b/wiki/entities/paper-loco-manip-161-005-chip.md index 5d2e35e18..9d134c691 100644 --- a/wiki/entities/paper-loco-manip-161-005-chip.md +++ b/wiki/entities/paper-loco-manip-161-005-chip.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "CHIP 把本体状态与关节序列、接触力/触觉信号转成可跟踪的身体目标，并通过扩散策略/流匹配、分层技能/专家策略训练或组合全身策略，最终输出全身轨迹/动作序列、末端执行器/腕手目标、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-006-clone.md b/wiki/entities/paper-loco-manip-161-006-clone.md index 90af3d16f..25778282b 100644 --- a/wiki/entities/paper-loco-manip-161-006-clone.md +++ b/wiki/entities/paper-loco-manip-161-006-clone.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "CLONE 主要解决数据闭环：用遥操作/外骨骼数据采集人类操作和机器人状态，再通过分层技能/专家策略、闭环纠错/人类干预转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-007-exbody2.md b/wiki/entities/paper-loco-manip-161-007-exbody2.md index 321f192df..04b731d14 100644 --- a/wiki/entities/paper-loco-manip-161-007-exbody2.md +++ b/wiki/entities/paper-loco-manip-161-007-exbody2.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "ExBody2 把本体状态与关节序列、人类视频/动捕轨迹转成可跟踪的身体目标，并通过异构动捕与合成平衡数据、教师-学生知识迁移、IK/动作重定向训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-008-from-w1.md b/wiki/entities/paper-loco-manip-161-008-from-w1.md index e7f89f2ca..87ee1c983 100644 --- a/wiki/entities/paper-loco-manip-161-008-from-w1.md +++ b/wiki/entities/paper-loco-manip-161-008-from-w1.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "FRoM-W1 把语言指令、本体状态与关节序列、仿真交互数据转成可跟踪的身体目标，并通过PPO/RL 策略训练、ACT/行为克隆模仿学习、VLM 语义规划/路由训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把示范轨迹压成可监督的动作预测问题，再通过动作 chunk 或闭环执行降低时序抖动。" related: diff --git a/wiki/entities/paper-loco-manip-161-009-gmt.md b/wiki/entities/paper-loco-manip-161-009-gmt.md index fe55976f9..dfc7e84eb 100644 --- a/wiki/entities/paper-loco-manip-161-009-gmt.md +++ b/wiki/entities/paper-loco-manip-161-009-gmt.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "GMT 把本体状态与关节序列、仿真交互数据转成可跟踪的身体目标，并通过扩散策略/流匹配、分层技能/专家策略训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-010-hover.md b/wiki/entities/paper-loco-manip-161-010-hover.md index 3fc25f95e..a8c35044f 100644 --- a/wiki/entities/paper-loco-manip-161-010-hover.md +++ b/wiki/entities/paper-loco-manip-161-010-hover.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HOVER 先从本体状态与关节序列恢复场景、目标或运动表征，再用教师-学生知识迁移、全身控制器/WBC/MPC、分层技能/专家策略生成全身轨迹/动作序列、低层控制器目标。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-011-holomotion.md b/wiki/entities/paper-loco-manip-161-011-holomotion.md index 5dcebc29a..108bfb3e6 100644 --- a/wiki/entities/paper-loco-manip-161-011-holomotion.md +++ b/wiki/entities/paper-loco-manip-161-011-holomotion.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HoloMotion 把人类视频/动捕轨迹转成可跟踪的身体目标，并通过异构动捕与合成平衡数据、扩散策略/流匹配、MM-DiT/Transformer 动作头训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-012-humanplus.md b/wiki/entities/paper-loco-manip-161-012-humanplus.md index 3726a4d99..e63bd719f 100644 --- a/wiki/entities/paper-loco-manip-161-012-humanplus.md +++ b/wiki/entities/paper-loco-manip-161-012-humanplus.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HumanPlus 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、人类视频/动捕轨迹采集人类操作和机器人状态，再通过PPO/RL 策略训练、ACT/行为克隆模仿学习、分层技能/专家策略转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-013-kungfubot2.md b/wiki/entities/paper-loco-manip-161-013-kungfubot2.md index df0aff644..e020e3b1d 100644 --- a/wiki/entities/paper-loco-manip-161-013-kungfubot2.md +++ b/wiki/entities/paper-loco-manip-161-013-kungfubot2.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "KungfuBot2 把本体状态与关节序列、人类视频/动捕轨迹、仿真交互数据转成可跟踪的身体目标，并通过教师-学生知识迁移、ACT/行为克隆模仿学习、IK/动作重定向训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-014-mosaic.md b/wiki/entities/paper-loco-manip-161-014-mosaic.md index 4e3433a9a..fcbfe2305 100644 --- a/wiki/entities/paper-loco-manip-161-014-mosaic.md +++ b/wiki/entities/paper-loco-manip-161-014-mosaic.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "MOSAIC 主要解决数据闭环：用本体状态与关节序列、遥操作/外骨骼数据、仿真交互数据采集人类操作和机器人状态，再通过扩散策略/流匹配转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-015-make-tracking-easy.md b/wiki/entities/paper-loco-manip-161-015-make-tracking-easy.md index dd244752a..d270a4d31 100644 --- a/wiki/entities/paper-loco-manip-161-015-make-tracking-easy.md +++ b/wiki/entities/paper-loco-manip-161-015-make-tracking-easy.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Make Tracking Easy 把本体状态与关节序列、接触力/触觉信号转成可跟踪的身体目标，并通过PPO/RL 策略训练、ACT/行为克隆模仿学习、扩散策略/流匹配训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标、地形/场景表征。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-016-omnih2o.md b/wiki/entities/paper-loco-manip-161-016-omnih2o.md index ef30c20bf..0a6f946b0 100644 --- a/wiki/entities/paper-loco-manip-161-016-omnih2o.md +++ b/wiki/entities/paper-loco-manip-161-016-omnih2o.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "OmniH2O 主要解决数据闭环：用语言指令、相机图像/多视角观测、人类视频/动捕轨迹采集人类操作和机器人状态，再通过教师-学生知识迁移、PPO/RL 策略训练、ACT/行为克隆模仿学习转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-017-omniretarget.md b/wiki/entities/paper-loco-manip-161-017-omniretarget.md index 70ea54eab..cf018dd20 100644 --- a/wiki/entities/paper-loco-manip-161-017-omniretarget.md +++ b/wiki/entities/paper-loco-manip-161-017-omniretarget.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "OmniRetarget 先从本体状态与关节序列、人类视频/动捕轨迹、接触力/触觉信号恢复场景、目标或运动表征，再用PPO/RL 策略训练、ACT/行为克隆模仿学习、IK/动作重定向生成全身轨迹/动作序列、低层控制器目标、地形/场景表征。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-018-retargeting.md b/wiki/entities/paper-loco-manip-161-018-retargeting.md index 022d31cf8..e9d4ada80 100644 --- a/wiki/entities/paper-loco-manip-161-018-retargeting.md +++ b/wiki/entities/paper-loco-manip-161-018-retargeting.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Retargeting 主要解决数据闭环：用人类视频/动捕轨迹、遥操作/外骨骼数据、接触力/触觉信号采集人类操作和机器人状态，再通过PPO/RL 策略训练、IK/动作重定向、分层技能/专家策略转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-019-sonic.md b/wiki/entities/paper-loco-manip-161-019-sonic.md deleted file mode 100644 index 52358b9ef..000000000 --- a/wiki/entities/paper-loco-manip-161-019-sonic.md +++ /dev/null @@ -1,83 +0,0 @@ ---- -type: entity -tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] -status: complete -updated: 2026-06-26 -venue: curated -summary: "SONIC 的实现路径是先把人类视频/动捕轨迹、遥操作/外骨骼数据编码成多模态表征，再用AMP/运动先验、VLA 多模态动作模型预测全身轨迹/动作序列、低层控制器目标。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" -related: - - ../overview/humanoid-loco-manip-161-papers-technology-map.md - - ../overview/loco-manip-161-category-01-motion-base-wbt.md - - ../tasks/loco-manipulation.md - - ../entities/paper-sonic.md -sources: - - ../../sources/papers/loco_manip_161_survey_019_sonic.md - - ../../sources/blogs/wechat_embodied_ai_lab_humanoid_loco_manip_161_survey.md - - ../../sources/papers/humanoid_loco_manip_161_catalog.md ---- - -# SONIC - -**SONIC** 收录于 [具身智能研究室 · 人形 Loco-Manip 161 篇长文](https://mp.weixin.qq.com/s/pACh9EhsISiyPGdiiR0C3A) **第 019/161** 篇，归类为 **01 运控基座与通用全身跟踪**。 - -## 一句话定义 - -SONIC 的实现路径是先把人类视频/动捕轨迹、遥操作/外骨骼数据编码成多模态表征，再用AMP/运动先验、VLA 多模态动作模型预测全身轨迹/动作序列、低层控制器目标。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。 - -## 英文缩写速查 - -| 缩写 | 英文全称 | 简要说明 | -|------|----------|----------| -| Loco-Manip | Loco-Manipulation | 行走与操作动力学耦合的全身任务 | -| WBC | Whole-Body Control | 协调全身关节满足多任务/约束的控制层 | -| VLA | Vision-Language-Action | 视觉-语言-动作多模态策略 | - -## 为什么重要 - -- SONIC 的实现路径是先把人类视频/动捕轨迹、遥操作/外骨骼数据编码成多模态表征，再用AMP/运动先验、VLA 多模态动作模型预测全身轨迹/动作序列、低层控制器目标。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。 -- 人形 Loco-Manip 161 篇 **#019/161** · 运控基座与通用全身跟踪。 - -## 核心信息（索引级） - -| 字段 | 内容 | -|------|------| -| 编号 | 019/161 | -| 分组 | 01 运控基座与通用全身跟踪 | -| 原文题目 | SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control | -| 机构 | NVIDIA | -| 发表日期 | 2026年5月21日 | -| 论文/项目 | https://nvlabs.github.io/GEAR-SONIC/ | - -## 核心机制（归纳） - -### 策展导读要点 - -SONIC 的实现路径是先把人类视频/动捕轨迹、遥操作/外骨骼数据编码成多模态表征，再用AMP/运动先验、VLA 多模态动作模型预测全身轨迹/动作序列、低层控制器目标。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。 - -## 评测与指标（索引级） - -- 本条目为 161 篇策展索引级摘录，**未搬运原文量化 benchmark 与实机指标**；评测口径与具体数值以原文 PDF / 项目页为准。 -- 评测原始出处：[原文 / 项目页](https://nvlabs.github.io/GEAR-SONIC/)（见上方「核心信息」表「论文/项目」一行）。 -- 横向评测对照请回到 [分类 hub](../overview/loco-manip-161-category-01-motion-base-wbt.md) 与 [技术地图](../overview/humanoid-loco-manip-161-papers-technology-map.md)。 - -## 常见误区 - -1. 161 篇策展条目提供 **地图坐标**；量化 benchmark 与实机指标以原文 PDF / 项目页为准。 -2. Loco-manip 单篇工作不自动解决 **底层 WBC 鲁棒性**；须与运控/接触控制对照。 - -## 与其他页面的关系 - -- 技术地图：[humanoid-loco-manip-161-papers-technology-map.md](../overview/humanoid-loco-manip-161-papers-technology-map.md) -- 分类 hub：[loco-manip-161-category-01-motion-base-wbt.md](../overview/loco-manip-161-category-01-motion-base-wbt.md) -- 原始 source：[loco_manip_161_survey_019_sonic.md](../../sources/papers/loco_manip_161_survey_019_sonic.md) - -## 参考来源 - -- [loco_manip_161_survey_019_sonic.md](../../sources/papers/loco_manip_161_survey_019_sonic.md) — 161 篇策展摘录 -- [humanoid_loco_manip_161_catalog.md](../../sources/papers/humanoid_loco_manip_161_catalog.md) -- [wechat_embodied_ai_lab_humanoid_loco_manip_161_survey.md](../../sources/blogs/wechat_embodied_ai_lab_humanoid_loco_manip_161_survey.md) - -## 推荐继续阅读 - -- [Loco-Manipulation 任务页](../tasks/loco-manipulation.md) -- 同题深读/既有实体：[paper-sonic](../entities/paper-sonic.md) diff --git a/wiki/entities/paper-loco-manip-161-020-twist2.md b/wiki/entities/paper-loco-manip-161-020-twist2.md index 5f7229c9f..a9a7c0de8 100644 --- a/wiki/entities/paper-loco-manip-161-020-twist2.md +++ b/wiki/entities/paper-loco-manip-161-020-twist2.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "TWIST2 的实现路径是先把相机图像/多视角观测、人类视频/动捕轨迹、遥操作/外骨骼数据编码成多模态表征，再用ACT/行为克隆模仿学习、VLA 多模态动作模型、全身控制器/WBC/MPC预测全身轨迹/动作序列、低层控制器目标。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-021-twist.md b/wiki/entities/paper-loco-manip-161-021-twist.md index 2d482658b..aa4a1f0b5 100644 --- a/wiki/entities/paper-loco-manip-161-021-twist.md +++ b/wiki/entities/paper-loco-manip-161-021-twist.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "TWIST 主要解决数据闭环：用人类视频/动捕轨迹、遥操作/外骨骼数据、接触力/触觉信号采集人类操作和机器人状态，再通过异构动捕与合成平衡数据、PPO/RL 策略训练、ACT/行为克隆模仿学习转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-022-textop.md b/wiki/entities/paper-loco-manip-161-022-textop.md index 10e5835ec..c66d88251 100644 --- a/wiki/entities/paper-loco-manip-161-022-textop.md +++ b/wiki/entities/paper-loco-manip-161-022-textop.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "TextOp 主要解决数据闭环：用语言指令、遥操作/外骨骼数据采集人类操作和机器人状态，再通过扩散策略/流匹配、全身控制器/WBC/MPC转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-023-uniact.md b/wiki/entities/paper-loco-manip-161-023-uniact.md index c682b081b..ef4a05ed6 100644 --- a/wiki/entities/paper-loco-manip-161-023-uniact.md +++ b/wiki/entities/paper-loco-manip-161-023-uniact.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "UniAct 把语言指令转成可跟踪的身体目标，并通过策略网络和控制模块训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把策略网络和控制模块放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-024-unitracker.md b/wiki/entities/paper-loco-manip-161-024-unitracker.md index 575e73c69..ee99fb9f6 100644 --- a/wiki/entities/paper-loco-manip-161-024-unitracker.md +++ b/wiki/entities/paper-loco-manip-161-024-unitracker.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "UniTracker 把相机图像/多视角观测、本体状态与关节序列、仿真交互数据转成可跟踪的身体目标，并通过教师-学生知识迁移、PPO/RL 策略训练、扩散策略/流匹配训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-025-n025.md b/wiki/entities/paper-loco-manip-161-025-n025.md index 2e898f397..737791f85 100644 --- a/wiki/entities/paper-loco-manip-161-025-n025.md +++ b/wiki/entities/paper-loco-manip-161-025-n025.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用人类视频/动捕轨迹、遥操作/外骨骼数据、接触力/触觉信号采集人类操作和机器人状态，再通过PPO/RL 策略训练、分层技能/专家策略转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-026-n026.md b/wiki/entities/paper-loco-manip-161-026-n026.md index 017647db7..22231e84b 100644 --- a/wiki/entities/paper-loco-manip-161-026-n026.md +++ b/wiki/entities/paper-loco-manip-161-026-n026.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作把本体状态与关节序列、人类视频/动捕轨迹、仿真交互数据转成可跟踪的身体目标，并通过异构动捕与合成平衡数据、PPO/RL 策略训练、ACT/行为克隆模仿学习训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把示范轨迹压成可监督的动作预测问题，再通过动作 chunk 或闭环执行降低时序抖动。" related: diff --git a/wiki/entities/paper-loco-manip-161-027-n027.md b/wiki/entities/paper-loco-manip-161-027-n027.md index b548b38b2..7380ebe5f 100644 --- a/wiki/entities/paper-loco-manip-161-027-n027.md +++ b/wiki/entities/paper-loco-manip-161-027-n027.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用本体状态与关节序列、遥操作/外骨骼数据、仿真交互数据采集人类操作和机器人状态，再通过教师-学生知识迁移、扩散策略/流匹配、全身控制器/WBC/MPC转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-028-n028.md b/wiki/entities/paper-loco-manip-161-028-n028.md index 78ece5e67..a1641ad6a 100644 --- a/wiki/entities/paper-loco-manip-161-028-n028.md +++ b/wiki/entities/paper-loco-manip-161-028-n028.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作把相机图像/多视角观测、本体状态与关节序列、仿真交互数据转成可跟踪的身体目标，并通过世界模型/视频预测、全身控制器/WBC/MPC训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把世界模型/视频预测、全身控制器/WBC/MPC放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-029-n029.md b/wiki/entities/paper-loco-manip-161-029-n029.md index 9c17a558f..0367e6f39 100644 --- a/wiki/entities/paper-loco-manip-161-029-n029.md +++ b/wiki/entities/paper-loco-manip-161-029-n029.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、人类视频/动捕轨迹采集人类操作和机器人状态，再通过IK/动作重定向转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是把IK/动作重定向放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-030-n030.md b/wiki/entities/paper-loco-manip-161-030-n030.md index a39920b08..d2f2d4cd1 100644 --- a/wiki/entities/paper-loco-manip-161-030-n030.md +++ b/wiki/entities/paper-loco-manip-161-030-n030.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从仿真交互数据、接触力/触觉信号恢复场景、目标或运动表征，再用扩散策略/流匹配生成全身轨迹/动作序列、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-031-n031.md b/wiki/entities/paper-loco-manip-161-031-n031.md index f151899de..6d79c7cd5 100644 --- a/wiki/entities/paper-loco-manip-161-031-n031.md +++ b/wiki/entities/paper-loco-manip-161-031-n031.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用本体状态与关节序列、人类视频/动捕轨迹、遥操作/外骨骼数据采集人类操作和机器人状态，再通过教师-学生知识迁移、ACT/行为克隆模仿学习、扩散策略/流匹配转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-032-amo.md b/wiki/entities/paper-loco-manip-161-032-amo.md index fa29088d5..db7e568ae 100644 --- a/wiki/entities/paper-loco-manip-161-032-amo.md +++ b/wiki/entities/paper-loco-manip-161-032-amo.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "AMO 把本体状态与关节序列、仿真交互数据、接触力/触觉信号转成可跟踪的身体目标，并通过PPO/RL 策略训练、ACT/行为克隆模仿学习、全身控制器/WBC/MPC训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把示范轨迹压成可监督的动作预测问题，再通过动作 chunk 或闭环执行降低时序抖动。" related: diff --git a/wiki/entities/paper-loco-manip-161-033-ceer.md b/wiki/entities/paper-loco-manip-161-033-ceer.md index 516ba68ad..a29d085b1 100644 --- a/wiki/entities/paper-loco-manip-161-033-ceer.md +++ b/wiki/entities/paper-loco-manip-161-033-ceer.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "CEER 主要解决数据闭环：用本体状态与关节序列、遥操作/外骨骼数据、仿真交互数据采集人类操作和机器人状态，再通过教师-学生知识迁移、全身控制器/WBC/MPC、分层技能/专家策略转成可训练、可复用的全身轨迹/动作序列、末端执行器/腕手目标、低层控制器目标。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-034-cola.md b/wiki/entities/paper-loco-manip-161-034-cola.md index 4e8553cd8..21d804eb1 100644 --- a/wiki/entities/paper-loco-manip-161-034-cola.md +++ b/wiki/entities/paper-loco-manip-161-034-cola.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "COLA 先从本体状态与关节序列、人类视频/动捕轨迹、仿真交互数据恢复场景、目标或运动表征，再用PPO/RL 策略训练、世界模型/视频预测、闭环纠错/人类干预生成全身轨迹/动作序列、低层控制器目标。关键点是把PPO/RL 策略训练、世界模型/视频预测放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-035-coordex.md b/wiki/entities/paper-loco-manip-161-035-coordex.md index efcc47d2d..e48834f95 100644 --- a/wiki/entities/paper-loco-manip-161-035-coordex.md +++ b/wiki/entities/paper-loco-manip-161-035-coordex.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "CoorDex 把本体状态与关节序列、接触力/触觉信号转成可跟踪的身体目标，并通过PPO/RL 策略训练、扩散策略/流匹配、MM-DiT/Transformer 动作头训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标、地形/场景表征。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-036-falcon.md b/wiki/entities/paper-loco-manip-161-036-falcon.md index e601d32e1..526156d17 100644 --- a/wiki/entities/paper-loco-manip-161-036-falcon.md +++ b/wiki/entities/paper-loco-manip-161-036-falcon.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "FALCON 主要解决数据闭环：用本体状态与关节序列、遥操作/外骨骼数据、接触力/触觉信号采集人类操作和机器人状态，再通过全身控制器/WBC/MPC转成可训练、可复用的关节位置/力矩命令、全身轨迹/动作序列、末端执行器/腕手目标。关键点是把全身控制器/WBC/MPC放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-037-handoff.md b/wiki/entities/paper-loco-manip-161-037-handoff.md index 42ca80c08..141ba7a12 100644 --- a/wiki/entities/paper-loco-manip-161-037-handoff.md +++ b/wiki/entities/paper-loco-manip-161-037-handoff.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HANDOFF 把本体状态与关节序列转成可跟踪的身体目标，并通过教师-学生知识迁移、扩散策略/流匹配、VLM 语义规划/路由训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-038-hex.md b/wiki/entities/paper-loco-manip-161-038-hex.md index 592c4d6cd..6bea2e75a 100644 --- a/wiki/entities/paper-loco-manip-161-038-hex.md +++ b/wiki/entities/paper-loco-manip-161-038-hex.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HEX 的实现路径是先把本体状态与关节序列编码成多模态表征，再用PPO/RL 策略训练、VLA 多模态动作模型、VLM 语义规划/路由预测全身轨迹/动作序列、低层控制器目标。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-039-hmc.md b/wiki/entities/paper-loco-manip-161-039-hmc.md index 7d39941ee..779aa8f7c 100644 --- a/wiki/entities/paper-loco-manip-161-039-hmc.md +++ b/wiki/entities/paper-loco-manip-161-039-hmc.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HMC 主要解决数据闭环：用遥操作/外骨骼数据、接触力/触觉信号采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、分层技能/专家策略转成可训练、可复用的关节位置/力矩命令、全身轨迹/动作序列、低层控制器目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-040-homie.md b/wiki/entities/paper-loco-manip-161-040-homie.md index 7826eb81b..b0975d112 100644 --- a/wiki/entities/paper-loco-manip-161-040-homie.md +++ b/wiki/entities/paper-loco-manip-161-040-homie.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HOMIE 主要解决数据闭环：用本体状态与关节序列、遥操作/外骨骼数据、仿真交互数据采集人类操作和机器人状态，再通过PPO/RL 策略训练、全身控制器/WBC/MPC转成可训练、可复用的全身轨迹/动作序列、低层控制器目标、地形/场景表征。关键点是把PPO/RL 策略训练、全身控制器/WBC/MPC放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-041-hiwet.md b/wiki/entities/paper-loco-manip-161-041-hiwet.md index d16cf3a12..df92e8930 100644 --- a/wiki/entities/paper-loco-manip-161-041-hiwet.md +++ b/wiki/entities/paper-loco-manip-161-041-hiwet.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HiWET 把本体状态与关节序列、仿真交互数据、接触力/触觉信号转成可跟踪的身体目标，并通过PPO/RL 策略训练、VLM 语义规划/路由、分层技能/专家策略训练或组合全身策略，最终输出全身轨迹/动作序列、末端执行器/腕手目标、低层控制器目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-042-hold.md b/wiki/entities/paper-loco-manip-161-042-hold.md index 0c6b22367..972d15ae9 100644 --- a/wiki/entities/paper-loco-manip-161-042-hold.md +++ b/wiki/entities/paper-loco-manip-161-042-hold.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Hold 先从相机图像/多视角观测恢复场景、目标或运动表征，再用ACT/行为克隆模仿学习生成全身轨迹/动作序列、末端执行器/腕手目标、低层控制器目标。关键点是把示范轨迹压成可监督的动作预测问题，再通过动作 chunk 或闭环执行降低时序抖动。" related: diff --git a/wiki/entities/paper-loco-manip-161-043-pilot.md b/wiki/entities/paper-loco-manip-161-043-pilot.md index 24953e644..2639867b6 100644 --- a/wiki/entities/paper-loco-manip-161-043-pilot.md +++ b/wiki/entities/paper-loco-manip-161-043-pilot.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "PILOT 先从本体状态与关节序列、仿真交互数据、接触力/触觉信号恢复场景、目标或运动表征，再用PPO/RL 策略训练、全身控制器/WBC/MPC、分层技能/专家策略生成全身轨迹/动作序列、低层控制器目标、地形/场景表征。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-044-rpl.md b/wiki/entities/paper-loco-manip-161-044-rpl.md index 78e6af52b..74525de03 100644 --- a/wiki/entities/paper-loco-manip-161-044-rpl.md +++ b/wiki/entities/paper-loco-manip-161-044-rpl.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "RPL 先从相机图像/多视角观测、本体状态与关节序列、深度/点云/高度图恢复场景、目标或运动表征，再用下视深度相机和 U-Net 高度图重建、教师-学生知识迁移、世界模型/视频预测生成全身轨迹/动作序列、低层控制器目标、地形/场景表征。关键点是把地形重建、步态相位和全身姿态放进同一个控制回路，而不是把感知和运控拆成松散串联。" related: diff --git a/wiki/entities/paper-loco-manip-161-045-splitadapter.md b/wiki/entities/paper-loco-manip-161-045-splitadapter.md index 5e0bfb9d5..365f80b76 100644 --- a/wiki/entities/paper-loco-manip-161-045-splitadapter.md +++ b/wiki/entities/paper-loco-manip-161-045-splitadapter.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "SplitAdapter 把仿真交互数据转成可跟踪的身体目标，并通过PPO/RL 策略训练、扩散策略/流匹配、全身控制器/WBC/MPC训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-046-sumo.md b/wiki/entities/paper-loco-manip-161-046-sumo.md index b2fae3060..9a976e790 100644 --- a/wiki/entities/paper-loco-manip-161-046-sumo.md +++ b/wiki/entities/paper-loco-manip-161-046-sumo.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Sumo 把本体状态与关节序列、仿真交互数据转成可跟踪的身体目标，并通过扩散策略/流匹配、全身控制器/WBC/MPC、分层技能/专家策略训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-047-thor.md b/wiki/entities/paper-loco-manip-161-047-thor.md index 1a88b1670..3687c23ce 100644 --- a/wiki/entities/paper-loco-manip-161-047-thor.md +++ b/wiki/entities/paper-loco-manip-161-047-thor.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Thor 把相机图像/多视角观测、本体状态与关节序列、接触力/触觉信号转成可跟踪的身体目标，并通过PPO/RL 策略训练、扩散策略/流匹配、全身控制器/WBC/MPC训练或组合全身策略，最终输出关节位置/力矩命令、全身轨迹/动作序列、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-048-ulc.md b/wiki/entities/paper-loco-manip-161-048-ulc.md index b481fb343..3fa18192f 100644 --- a/wiki/entities/paper-loco-manip-161-048-ulc.md +++ b/wiki/entities/paper-loco-manip-161-048-ulc.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "ULC 把本体状态与关节序列、接触力/触觉信号转成可跟踪的身体目标，并通过PPO/RL 策略训练、全身控制器/WBC/MPC、分层技能/专家策略训练或组合全身策略，最终输出关节位置/力矩命令、全身轨迹/动作序列、低层控制器目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-049-viral.md b/wiki/entities/paper-loco-manip-161-049-viral.md index dd186f295..a6319485c 100644 --- a/wiki/entities/paper-loco-manip-161-049-viral.md +++ b/wiki/entities/paper-loco-manip-161-049-viral.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "VIRAL 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过教师-学生知识迁移、ACT/行为克隆模仿学习、分层技能/专家策略转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-050-vofa.md b/wiki/entities/paper-loco-manip-161-050-vofa.md index 75776f188..16942d790 100644 --- a/wiki/entities/paper-loco-manip-161-050-vofa.md +++ b/wiki/entities/paper-loco-manip-161-050-vofa.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "VOFA 把相机图像/多视角观测、本体状态与关节序列、仿真交互数据转成可跟踪的身体目标，并通过扩散策略/流匹配、全身控制器/WBC/MPC、分层技能/专家策略训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标、导航/到达目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-051-wholebodyvla.md b/wiki/entities/paper-loco-manip-161-051-wholebodyvla.md index a7f01e697..9b6f3239e 100644 --- a/wiki/entities/paper-loco-manip-161-051-wholebodyvla.md +++ b/wiki/entities/paper-loco-manip-161-051-wholebodyvla.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "WholeBodyVLA 的实现路径是先把相机图像/多视角观测、遥操作/外骨骼数据编码成多模态表征，再用VLA 多模态动作模型预测全身轨迹/动作序列、低层控制器目标。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-052-n052.md b/wiki/entities/paper-loco-manip-161-052-n052.md index 3beb36a56..ea9976bd5 100644 --- a/wiki/entities/paper-loco-manip-161-052-n052.md +++ b/wiki/entities/paper-loco-manip-161-052-n052.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作把相机图像/多视角观测、本体状态与关节序列、仿真交互数据转成可跟踪的身体目标，并通过PPO/RL 策略训练、扩散策略/流匹配训练或组合全身策略，最终输出全身轨迹/动作序列、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-053-n053.md b/wiki/entities/paper-loco-manip-161-053-n053.md index e7e62c106..8a9b87e41 100644 --- a/wiki/entities/paper-loco-manip-161-053-n053.md +++ b/wiki/entities/paper-loco-manip-161-053-n053.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过教师-学生知识迁移、全身控制器/WBC/MPC、闭环纠错/人类干预转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-054-n054.md b/wiki/entities/paper-loco-manip-161-054-n054.md index de8829239..4b7b15e73 100644 --- a/wiki/entities/paper-loco-manip-161-054-n054.md +++ b/wiki/entities/paper-loco-manip-161-054-n054.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用本体状态与关节序列、遥操作/外骨骼数据、仿真交互数据采集人类操作和机器人状态，再通过扩散策略/流匹配、全身控制器/WBC/MPC转成可训练、可复用的关节位置/力矩命令、全身轨迹/动作序列、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-055-n055.md b/wiki/entities/paper-loco-manip-161-055-n055.md index 552ed2466..40d754dbc 100644 --- a/wiki/entities/paper-loco-manip-161-055-n055.md +++ b/wiki/entities/paper-loco-manip-161-055-n055.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过AMP/运动先验、全身控制器/WBC/MPC转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是把AMP/运动先验、全身控制器/WBC/MPC放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-056-agile.md b/wiki/entities/paper-loco-manip-161-056-agile.md index c20ea3f05..e2f3b2dc8 100644 --- a/wiki/entities/paper-loco-manip-161-056-agile.md +++ b/wiki/entities/paper-loco-manip-161-056-agile.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "AGILE 先从相机图像/多视角观测、本体状态与关节序列、仿真交互数据恢复场景、目标或运动表征，再用PPO/RL 策略训练、全身控制器/WBC/MPC、分层技能/专家策略生成低层控制器目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-057-being-0.md b/wiki/entities/paper-loco-manip-161-057-being-0.md index 2a428cb09..4ff430749 100644 --- a/wiki/entities/paper-loco-manip-161-057-being-0.md +++ b/wiki/entities/paper-loco-manip-161-057-being-0.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Being-0 先从语言指令、相机图像/多视角观测恢复场景、目标或运动表征，再用VLM 语义规划/路由、分层技能/专家策略生成可执行动作命令。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-058-cybo-waiter.md b/wiki/entities/paper-loco-manip-161-058-cybo-waiter.md index 736025201..028cd686f 100644 --- a/wiki/entities/paper-loco-manip-161-058-cybo-waiter.md +++ b/wiki/entities/paper-loco-manip-161-058-cybo-waiter.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Cybo-Waiter 先从语言指令、相机图像/多视角观测恢复场景、目标或运动表征，再用扩散策略/流匹配、VLM 语义规划/路由、闭环纠错/人类干预生成全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-059-demohlm.md b/wiki/entities/paper-loco-manip-161-059-demohlm.md index 7085ee5a1..b20833a5c 100644 --- a/wiki/entities/paper-loco-manip-161-059-demohlm.md +++ b/wiki/entities/paper-loco-manip-161-059-demohlm.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "DemoHLM 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、全身控制器/WBC/MPC、闭环纠错/人类干预转成可训练、可复用的关节位置/力矩命令、全身轨迹/动作序列、低层控制器目标。关键点是把示范轨迹压成可监督的动作预测问题，再通过动作 chunk 或闭环执行降低时序抖动。" related: diff --git a/wiki/entities/paper-loco-manip-161-060-egohumanoid.md b/wiki/entities/paper-loco-manip-161-060-egohumanoid.md index c19291866..e305a0198 100644 --- a/wiki/entities/paper-loco-manip-161-060-egohumanoid.md +++ b/wiki/entities/paper-loco-manip-161-060-egohumanoid.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "EgoHumanoid 的实现路径是先把相机图像/多视角观测、人类视频/动捕轨迹、遥操作/外骨骼数据编码成多模态表征，再用ACT/行为克隆模仿学习、VLA 多模态动作模型预测地形/场景表征。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-061-grail.md b/wiki/entities/paper-loco-manip-161-061-grail.md index c786789c6..4e0eba06d 100644 --- a/wiki/entities/paper-loco-manip-161-061-grail.md +++ b/wiki/entities/paper-loco-manip-161-061-grail.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "GRAIL 先从相机图像/多视角观测、本体状态与关节序列、人类视频/动捕轨迹恢复场景、目标或运动表征，再用ACT/行为克隆模仿学习、扩散策略/流匹配、世界模型/视频预测生成末端执行器/腕手目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-062-h2-compact.md b/wiki/entities/paper-loco-manip-161-062-h2-compact.md index 026326d35..cc9c7f5aa 100644 --- a/wiki/entities/paper-loco-manip-161-062-h2-compact.md +++ b/wiki/entities/paper-loco-manip-161-062-h2-compact.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "H2-COMPACT 先从相机图像/多视角观测、本体状态与关节序列、仿真交互数据恢复场景、目标或运动表征，再用扩散策略/流匹配、分层技能/专家策略生成关节位置/力矩命令、全身轨迹/动作序列、地形/场景表征。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-063-halomi.md b/wiki/entities/paper-loco-manip-161-063-halomi.md index 8a2f40174..eaddee772 100644 --- a/wiki/entities/paper-loco-manip-161-063-halomi.md +++ b/wiki/entities/paper-loco-manip-161-063-halomi.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HALOMI 先从相机图像/多视角观测恢复场景、目标或运动表征，再用ACT/行为克隆模仿学习、扩散策略/流匹配生成关节位置/力矩命令、全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-064-hmc.md b/wiki/entities/paper-loco-manip-161-064-hmc.md index f5bc65170..cfe0bca28 100644 --- a/wiki/entities/paper-loco-manip-161-064-hmc.md +++ b/wiki/entities/paper-loco-manip-161-064-hmc.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HMC 主要解决数据闭环：用相机图像/多视角观测、遥操作/外骨骼数据、接触力/触觉信号采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、分层技能/专家策略转成可训练、可复用的关节位置/力矩命令。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-065-hypermotion.md b/wiki/entities/paper-loco-manip-161-065-hypermotion.md index 63a2fc9f4..0a366debd 100644 --- a/wiki/entities/paper-loco-manip-161-065-hypermotion.md +++ b/wiki/entities/paper-loco-manip-161-065-hypermotion.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HYPERmotion 先从语言指令、相机图像/多视角观测、本体状态与关节序列恢复场景、目标或运动表征，再用PPO/RL 策略训练、VLM 语义规划/路由、全身控制器/WBC/MPC生成可执行动作命令。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-066-hiwet.md b/wiki/entities/paper-loco-manip-161-066-hiwet.md index a8d19a1f0..94c487e08 100644 --- a/wiki/entities/paper-loco-manip-161-066-hiwet.md +++ b/wiki/entities/paper-loco-manip-161-066-hiwet.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HiWET 先从相机图像/多视角观测、本体状态与关节序列、仿真交互数据恢复场景、目标或运动表征，再用PPO/RL 策略训练、VLM 语义规划/路由、分层技能/专家策略生成末端执行器/腕手目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-067-humanoidexo.md b/wiki/entities/paper-loco-manip-161-067-humanoidexo.md index 5e14bed27..14a7ea79c 100644 --- a/wiki/entities/paper-loco-manip-161-067-humanoidexo.md +++ b/wiki/entities/paper-loco-manip-161-067-humanoidexo.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HumanoidExo 主要解决数据闭环：用相机图像/多视角观测、人类视频/动捕轨迹、遥操作/外骨骼数据采集人类操作和机器人状态，再通过扩散策略/流匹配、全身控制器/WBC/MPC、分层技能/专家策略转成可训练、可复用的全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-068-humanoidmimicgen.md b/wiki/entities/paper-loco-manip-161-068-humanoidmimicgen.md index da5a2b496..418845644 100644 --- a/wiki/entities/paper-loco-manip-161-068-humanoidmimicgen.md +++ b/wiki/entities/paper-loco-manip-161-068-humanoidmimicgen.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HumanoidMimicGen 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、分层技能/专家策略转成可训练、可复用的全身轨迹/动作序列。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-069-humanoid.md b/wiki/entities/paper-loco-manip-161-069-humanoid.md index 8e19525be..fa32b37f3 100644 --- a/wiki/entities/paper-loco-manip-161-069-humanoid.md +++ b/wiki/entities/paper-loco-manip-161-069-humanoid.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Humanoid 主要解决数据闭环：用语言指令、相机图像/多视角观测、遥操作/外骨骼数据采集人类操作和机器人状态，再通过扩散策略/流匹配转成可训练、可复用的可执行动作命令。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-070-oasis.md b/wiki/entities/paper-loco-manip-161-070-oasis.md index b5e81fcc9..678a25b8f 100644 --- a/wiki/entities/paper-loco-manip-161-070-oasis.md +++ b/wiki/entities/paper-loco-manip-161-070-oasis.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "OASIS 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、分层技能/专家策略转成可训练、可复用的可执行动作命令。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-071-omnidp.md b/wiki/entities/paper-loco-manip-161-071-omnidp.md index f3ada61c0..9da78e3a8 100644 --- a/wiki/entities/paper-loco-manip-161-071-omnidp.md +++ b/wiki/entities/paper-loco-manip-161-071-omnidp.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "OmniDP 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过策略网络和控制模块转成可训练、可复用的可执行动作命令。关键点是把策略网络和控制模块放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-072-opt2skill.md b/wiki/entities/paper-loco-manip-161-072-opt2skill.md index 6dc2bd191..32c0895f6 100644 --- a/wiki/entities/paper-loco-manip-161-072-opt2skill.md +++ b/wiki/entities/paper-loco-manip-161-072-opt2skill.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Opt2Skill 先从相机图像/多视角观测、仿真交互数据、接触力/触觉信号恢复场景、目标或运动表征，再用PPO/RL 策略训练、ACT/行为克隆模仿学习、IK/动作重定向生成关节位置/力矩命令、全身轨迹/动作序列。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-073-physhsi.md b/wiki/entities/paper-loco-manip-161-073-physhsi.md index 3c3166df0..b6abc0548 100644 --- a/wiki/entities/paper-loco-manip-161-073-physhsi.md +++ b/wiki/entities/paper-loco-manip-161-073-physhsi.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "PhysHSI 先从相机图像/多视角观测、本体状态与关节序列、深度/点云/高度图恢复场景、目标或运动表征，再用AMP/运动先验、DINO/视觉特征抽取生成地形/场景表征。关键点是把AMP/运动先验、DINO/视觉特征抽取放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-074-pro-hoi.md b/wiki/entities/paper-loco-manip-161-074-pro-hoi.md index 9a119772a..b3bbe85ca 100644 --- a/wiki/entities/paper-loco-manip-161-074-pro-hoi.md +++ b/wiki/entities/paper-loco-manip-161-074-pro-hoi.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Pro-HOI 先从相机图像/多视角观测、本体状态与关节序列、人类视频/动捕轨迹恢复场景、目标或运动表征，再用扩散策略/流匹配、IK/动作重定向、全身控制器/WBC/MPC生成全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-075-simple.md b/wiki/entities/paper-loco-manip-161-075-simple.md index 766aeb0cc..0b1e16b74 100644 --- a/wiki/entities/paper-loco-manip-161-075-simple.md +++ b/wiki/entities/paper-loco-manip-161-075-simple.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "SIMPLE 的实现路径是先把相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据编码成多模态表征，再用VLA 多模态动作模型、世界模型/视频预测预测全身轨迹/动作序列。关键点是让视频/世界模型提供可预测的物理先验，再由动作头把语义目标变成连续控制。" related: diff --git a/wiki/entities/paper-loco-manip-161-076-sugar.md b/wiki/entities/paper-loco-manip-161-076-sugar.md index ceddc86e2..d2f983300 100644 --- a/wiki/entities/paper-loco-manip-161-076-sugar.md +++ b/wiki/entities/paper-loco-manip-161-076-sugar.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "SUGAR 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、人类视频/动捕轨迹采集人类操作和机器人状态，再通过AMP/运动先验、ACT/行为克隆模仿学习、扩散策略/流匹配转成可训练、可复用的可执行动作命令。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-077-skillblender.md b/wiki/entities/paper-loco-manip-161-077-skillblender.md index 81e864f22..5ed39a934 100644 --- a/wiki/entities/paper-loco-manip-161-077-skillblender.md +++ b/wiki/entities/paper-loco-manip-161-077-skillblender.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "SkillBlender 先从相机图像/多视角观测、本体状态与关节序列、接触力/触觉信号恢复场景、目标或运动表征，再用PPO/RL 策略训练、扩散策略/流匹配、全身控制器/WBC/MPC生成全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-078-stageact.md b/wiki/entities/paper-loco-manip-161-078-stageact.md index 340c4d9ad..778c5b63c 100644 --- a/wiki/entities/paper-loco-manip-161-078-stageact.md +++ b/wiki/entities/paper-loco-manip-161-078-stageact.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "StageACT 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、扩散策略/流匹配、VLM 语义规划/路由转成可训练、可复用的可执行动作命令。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-079-synagent.md b/wiki/entities/paper-loco-manip-161-079-synagent.md index cf5bd3402..292708814 100644 --- a/wiki/entities/paper-loco-manip-161-079-synagent.md +++ b/wiki/entities/paper-loco-manip-161-079-synagent.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "SynAgent 先从相机图像/多视角观测恢复场景、目标或运动表征，再用教师-学生知识迁移、PPO/RL 策略训练、扩散策略/流匹配生成全身轨迹/动作序列、地形/场景表征。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-080-trajbooster.md b/wiki/entities/paper-loco-manip-161-080-trajbooster.md index bb7f1a6ac..58f0b89c5 100644 --- a/wiki/entities/paper-loco-manip-161-080-trajbooster.md +++ b/wiki/entities/paper-loco-manip-161-080-trajbooster.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "TrajBooster 的实现路径是先把相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据编码成多模态表征，再用ACT/行为克隆模仿学习、VLA 多模态动作模型、IK/动作重定向预测全身轨迹/动作序列、末端执行器/腕手目标。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-081-viral.md b/wiki/entities/paper-loco-manip-161-081-viral.md index 982013f39..0d1dda2c7 100644 --- a/wiki/entities/paper-loco-manip-161-081-viral.md +++ b/wiki/entities/paper-loco-manip-161-081-viral.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "VIRAL 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过教师-学生知识迁移、ACT/行为克隆模仿学习、分层技能/专家策略转成可训练、可复用的可执行动作命令。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-082-visualmimic.md b/wiki/entities/paper-loco-manip-161-082-visualmimic.md index 134601252..bf46ac0b7 100644 --- a/wiki/entities/paper-loco-manip-161-082-visualmimic.md +++ b/wiki/entities/paper-loco-manip-161-082-visualmimic.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "VisualMimic 先从相机图像/多视角观测、本体状态与关节序列、人类视频/动捕轨迹恢复场景、目标或运动表征，再用教师-学生知识迁移、全身控制器/WBC/MPC生成可执行动作命令。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-083-wholebodyvla.md b/wiki/entities/paper-loco-manip-161-083-wholebodyvla.md index cd9768a57..dd748ec19 100644 --- a/wiki/entities/paper-loco-manip-161-083-wholebodyvla.md +++ b/wiki/entities/paper-loco-manip-161-083-wholebodyvla.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "WholeBodyVLA 的实现路径是先把相机图像/多视角观测、遥操作/外骨骼数据编码成多模态表征，再用VLA 多模态动作模型预测全身轨迹/动作序列。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-084-n084.md b/wiki/entities/paper-loco-manip-161-084-n084.md index 83efbbd08..dfeacfd6e 100644 --- a/wiki/entities/paper-loco-manip-161-084-n084.md +++ b/wiki/entities/paper-loco-manip-161-084-n084.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用相机图像/多视角观测、人类视频/动捕轨迹、遥操作/外骨骼数据采集人类操作和机器人状态，再通过PPO/RL 策略训练、分层技能/专家策略转成可训练、可复用的全身轨迹/动作序列。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-085-n085.md b/wiki/entities/paper-loco-manip-161-085-n085.md index 6e48f1a80..7a9319820 100644 --- a/wiki/entities/paper-loco-manip-161-085-n085.md +++ b/wiki/entities/paper-loco-manip-161-085-n085.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、本体状态与关节序列、接触力/触觉信号恢复场景、目标或运动表征，再用ACT/行为克隆模仿学习、扩散策略/流匹配、全身控制器/WBC/MPC生成可执行动作命令。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-086-n086.md b/wiki/entities/paper-loco-manip-161-086-n086.md index af3671c34..d276f62d1 100644 --- a/wiki/entities/paper-loco-manip-161-086-n086.md +++ b/wiki/entities/paper-loco-manip-161-086-n086.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "使用3D扩散策略的通用人形操作 主要解决数据闭环：用相机图像/多视角观测、遥操作/外骨骼数据、深度/点云/高度图采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、扩散策略/流匹配、分层技能/专家策略转成可训练、可复用的可执行动作命令。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-087-n087.md b/wiki/entities/paper-loco-manip-161-087-n087.md index b77325a2e..ce1c01901 100644 --- a/wiki/entities/paper-loco-manip-161-087-n087.md +++ b/wiki/entities/paper-loco-manip-161-087-n087.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用相机图像/多视角观测、遥操作/外骨骼数据采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、扩散策略/流匹配转成可训练、可复用的可执行动作命令。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-088-n088.md b/wiki/entities/paper-loco-manip-161-088-n088.md index 5ca924489..7eda3886e 100644 --- a/wiki/entities/paper-loco-manip-161-088-n088.md +++ b/wiki/entities/paper-loco-manip-161-088-n088.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、本体状态与关节序列、深度/点云/高度图恢复场景、目标或运动表征，再用扩散策略/流匹配、IK/动作重定向生成关节位置/力矩命令、地形/场景表征。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-089-n089.md b/wiki/entities/paper-loco-manip-161-089-n089.md index 03bfa5bde..e02b8fcc6 100644 --- a/wiki/entities/paper-loco-manip-161-089-n089.md +++ b/wiki/entities/paper-loco-manip-161-089-n089.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过教师-学生知识迁移、全身控制器/WBC/MPC、闭环纠错/人类干预转成可训练、可复用的可执行动作命令。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-090-n090.md b/wiki/entities/paper-loco-manip-161-090-n090.md index 3749d724d..899bacd6b 100644 --- a/wiki/entities/paper-loco-manip-161-090-n090.md +++ b/wiki/entities/paper-loco-manip-161-090-n090.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、本体状态与关节序列、深度/点云/高度图恢复场景、目标或运动表征，再用PPO/RL 策略训练、扩散策略/流匹配、世界模型/视频预测生成低层控制器目标、地形/场景表征。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-091-n091.md b/wiki/entities/paper-loco-manip-161-091-n091.md index dced65537..cdbee35c3 100644 --- a/wiki/entities/paper-loco-manip-161-091-n091.md +++ b/wiki/entities/paper-loco-manip-161-091-n091.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用相机图像/多视角观测、遥操作/外骨骼数据、仿真交互数据采集人类操作和机器人状态，再通过扩散策略/流匹配、全身控制器/WBC/MPC、分层技能/专家策略转成可训练、可复用的低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-092-n092.md b/wiki/entities/paper-loco-manip-161-092-n092.md index 871a739f2..0a7a30204 100644 --- a/wiki/entities/paper-loco-manip-161-092-n092.md +++ b/wiki/entities/paper-loco-manip-161-092-n092.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过PPO/RL 策略训练、扩散策略/流匹配、MM-DiT/Transformer 动作头转成可训练、可复用的末端执行器/腕手目标、动作 chunk/token。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-093-dreamcontrol.md b/wiki/entities/paper-loco-manip-161-093-dreamcontrol.md index a0dc40f93..a072f604c 100644 --- a/wiki/entities/paper-loco-manip-161-093-dreamcontrol.md +++ b/wiki/entities/paper-loco-manip-161-093-dreamcontrol.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "DreamControl 先从本体状态与关节序列、人类视频/动捕轨迹、仿真交互数据恢复场景、目标或运动表征，再用PPO/RL 策略训练、扩散策略/流匹配、分层技能/专家策略生成全身轨迹/动作序列、地形/场景表征。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-094-emotion.md b/wiki/entities/paper-loco-manip-161-094-emotion.md index 4392de3c9..83858cfbe 100644 --- a/wiki/entities/paper-loco-manip-161-094-emotion.md +++ b/wiki/entities/paper-loco-manip-161-094-emotion.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "EMOTION 先从视觉、状态和动作数据恢复场景、目标或运动表征，再用策略网络和控制模块生成全身轨迹/动作序列。关键点是把策略网络和控制模块放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-095-egoprimo.md b/wiki/entities/paper-loco-manip-161-095-egoprimo.md index 0beb656b8..378275c8c 100644 --- a/wiki/entities/paper-loco-manip-161-095-egoprimo.md +++ b/wiki/entities/paper-loco-manip-161-095-egoprimo.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "EgoPriMo 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用AMP/运动先验、ACT/行为克隆模仿学习、扩散策略/流匹配预测全身轨迹/动作序列。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-096-from-w1.md b/wiki/entities/paper-loco-manip-161-096-from-w1.md index 89dda16d8..3d2abf410 100644 --- a/wiki/entities/paper-loco-manip-161-096-from-w1.md +++ b/wiki/entities/paper-loco-manip-161-096-from-w1.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "FRoM-W1 先从语言指令、本体状态与关节序列、仿真交互数据恢复场景、目标或运动表征，再用PPO/RL 策略训练、ACT/行为克隆模仿学习、VLM 语义规划/路由生成全身轨迹/动作序列。关键点是把示范轨迹压成可监督的动作预测问题，再通过动作 chunk 或闭环执行降低时序抖动。" related: diff --git a/wiki/entities/paper-loco-manip-161-097-harmon.md b/wiki/entities/paper-loco-manip-161-097-harmon.md index cb2000404..c890a4029 100644 --- a/wiki/entities/paper-loco-manip-161-097-harmon.md +++ b/wiki/entities/paper-loco-manip-161-097-harmon.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Harmon 先从语言指令、本体状态与关节序列、人类视频/动捕轨迹恢复场景、目标或运动表征，再用AMP/运动先验、扩散策略/流匹配、VLM 语义规划/路由生成全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-098-langwbc.md b/wiki/entities/paper-loco-manip-161-098-langwbc.md index 3b916527b..91072c3a7 100644 --- a/wiki/entities/paper-loco-manip-161-098-langwbc.md +++ b/wiki/entities/paper-loco-manip-161-098-langwbc.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "LangWBC 先从语言指令、本体状态与关节序列、仿真交互数据恢复场景、目标或运动表征，再用教师-学生知识迁移、PPO/RL 策略训练、扩散策略/流匹配生成全身轨迹/动作序列、低层控制器目标。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-099-motiondisco.md b/wiki/entities/paper-loco-manip-161-099-motiondisco.md index c0360aa55..da06ed852 100644 --- a/wiki/entities/paper-loco-manip-161-099-motiondisco.md +++ b/wiki/entities/paper-loco-manip-161-099-motiondisco.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "MotionDisco 主要解决数据闭环：用遥操作/外骨骼数据、接触力/触觉信号采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、IK/动作重定向、分层技能/专家策略转成可训练、可复用的全身轨迹/动作序列。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-100-motionwam.md b/wiki/entities/paper-loco-manip-161-100-motionwam.md index 78922649a..d5fac38bd 100644 --- a/wiki/entities/paper-loco-manip-161-100-motionwam.md +++ b/wiki/entities/paper-loco-manip-161-100-motionwam.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "MotionWAM 的实现路径是先把相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用ACT/行为克隆模仿学习、扩散策略/流匹配、VLA 多模态动作模型预测全身轨迹/动作序列、低层控制器目标、地形/场景表征。关键点是让视频/世界模型提供可预测的物理先验，再由动作头把语义目标变成连续控制。" related: diff --git a/wiki/entities/paper-loco-manip-161-101-omg.md b/wiki/entities/paper-loco-manip-161-101-omg.md index 4ab92f280..5af83bfe2 100644 --- a/wiki/entities/paper-loco-manip-161-101-omg.md +++ b/wiki/entities/paper-loco-manip-161-101-omg.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "OMG 先从本体状态与关节序列恢复场景、目标或运动表征，再用扩散策略/流匹配、VLM 语义规划/路由、全身控制器/WBC/MPC生成全身轨迹/动作序列、低层控制器目标、地形/场景表征。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-102-robot-motion-diffusion-model.md b/wiki/entities/paper-loco-manip-161-102-robot-motion-diffusion-model.md index 83fdad21b..5e83422ca 100644 --- a/wiki/entities/paper-loco-manip-161-102-robot-motion-diffusion-model.md +++ b/wiki/entities/paper-loco-manip-161-102-robot-motion-diffusion-model.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Robot Motion Diffusion Model 先从本体状态与关节序列、仿真交互数据、接触力/触觉信号恢复场景、目标或运动表征，再用PPO/RL 策略训练、ACT/行为克隆模仿学习、扩散策略/流匹配生成全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-103-sonic.md b/wiki/entities/paper-loco-manip-161-103-sonic.md deleted file mode 100644 index 5b63dc982..000000000 --- a/wiki/entities/paper-loco-manip-161-103-sonic.md +++ /dev/null @@ -1,83 +0,0 @@ ---- -type: entity -tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] -status: complete -updated: 2026-06-26 -venue: curated -summary: "SONIC 的实现路径是先把人类视频/动捕轨迹、遥操作/外骨骼数据编码成多模态表征，再用AMP/运动先验、VLA 多模态动作模型预测全身轨迹/动作序列。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" -related: - - ../overview/humanoid-loco-manip-161-papers-technology-map.md - - ../overview/loco-manip-161-category-04-generative-language-trajectory.md - - ../tasks/loco-manipulation.md - - ../entities/paper-sonic.md -sources: - - ../../sources/papers/loco_manip_161_survey_103_sonic.md - - ../../sources/blogs/wechat_embodied_ai_lab_humanoid_loco_manip_161_survey.md - - ../../sources/papers/humanoid_loco_manip_161_catalog.md ---- - -# SONIC - -**SONIC** 收录于 [具身智能研究室 · 人形 Loco-Manip 161 篇长文](https://mp.weixin.qq.com/s/pACh9EhsISiyPGdiiR0C3A) **第 103/161** 篇，归类为 **04 生成式运动、语言控制与轨迹规划**。 - -## 一句话定义 - -SONIC 的实现路径是先把人类视频/动捕轨迹、遥操作/外骨骼数据编码成多模态表征，再用AMP/运动先验、VLA 多模态动作模型预测全身轨迹/动作序列。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。 - -## 英文缩写速查 - -| 缩写 | 英文全称 | 简要说明 | -|------|----------|----------| -| Loco-Manip | Loco-Manipulation | 行走与操作动力学耦合的全身任务 | -| WBC | Whole-Body Control | 协调全身关节满足多任务/约束的控制层 | -| VLA | Vision-Language-Action | 视觉-语言-动作多模态策略 | - -## 为什么重要 - -- SONIC 的实现路径是先把人类视频/动捕轨迹、遥操作/外骨骼数据编码成多模态表征，再用AMP/运动先验、VLA 多模态动作模型预测全身轨迹/动作序列。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。 -- 人形 Loco-Manip 161 篇 **#103/161** · 生成式运动、语言控制与轨迹规划。 - -## 核心信息（索引级） - -| 字段 | 内容 | -|------|------| -| 编号 | 103/161 | -| 分组 | 04 生成式运动、语言控制与轨迹规划 | -| 原文题目 | SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control | -| 机构 | NVIDIA | -| 发表日期 | 2026年5月21日 | -| 论文/项目 | https://nvlabs.github.io/GEAR-SONIC/ | - -## 核心机制（归纳） - -### 策展导读要点 - -SONIC 的实现路径是先把人类视频/动捕轨迹、遥操作/外骨骼数据编码成多模态表征，再用AMP/运动先验、VLA 多模态动作模型预测全身轨迹/动作序列。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。 - -## 评测与指标（索引级） - -- 本条目为 161 篇策展索引级摘录，**未搬运原文量化 benchmark 与实机指标**；评测口径与具体数值以原文 PDF / 项目页为准。 -- 评测原始出处：[原文 / 项目页](https://nvlabs.github.io/GEAR-SONIC/)（见上方「核心信息」表「论文/项目」一行）。 -- 横向评测对照请回到 [分类 hub](../overview/loco-manip-161-category-04-generative-language-trajectory.md) 与 [技术地图](../overview/humanoid-loco-manip-161-papers-technology-map.md)。 - -## 常见误区 - -1. 161 篇策展条目提供 **地图坐标**；量化 benchmark 与实机指标以原文 PDF / 项目页为准。 -2. Loco-manip 单篇工作不自动解决 **底层 WBC 鲁棒性**；须与运控/接触控制对照。 - -## 与其他页面的关系 - -- 技术地图：[humanoid-loco-manip-161-papers-technology-map.md](../overview/humanoid-loco-manip-161-papers-technology-map.md) -- 分类 hub：[loco-manip-161-category-04-generative-language-trajectory.md](../overview/loco-manip-161-category-04-generative-language-trajectory.md) -- 原始 source：[loco_manip_161_survey_103_sonic.md](../../sources/papers/loco_manip_161_survey_103_sonic.md) - -## 参考来源 - -- [loco_manip_161_survey_103_sonic.md](../../sources/papers/loco_manip_161_survey_103_sonic.md) — 161 篇策展摘录 -- [humanoid_loco_manip_161_catalog.md](../../sources/papers/humanoid_loco_manip_161_catalog.md) -- [wechat_embodied_ai_lab_humanoid_loco_manip_161_survey.md](../../sources/blogs/wechat_embodied_ai_lab_humanoid_loco_manip_161_survey.md) - -## 推荐继续阅读 - -- [Loco-Manipulation 任务页](../tasks/loco-manipulation.md) -- 同题深读/既有实体：[paper-sonic](../entities/paper-sonic.md) diff --git a/wiki/entities/paper-loco-manip-161-104-safeflow.md b/wiki/entities/paper-loco-manip-161-104-safeflow.md index 8bfd8ceaf..9014dc29c 100644 --- a/wiki/entities/paper-loco-manip-161-104-safeflow.md +++ b/wiki/entities/paper-loco-manip-161-104-safeflow.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "SafeFlow 先从本体状态与关节序列恢复场景、目标或运动表征，再用扩散策略/流匹配、全身控制器/WBC/MPC、分层技能/专家策略生成全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-105-textop.md b/wiki/entities/paper-loco-manip-161-105-textop.md index e5cd53f7e..191ca77c0 100644 --- a/wiki/entities/paper-loco-manip-161-105-textop.md +++ b/wiki/entities/paper-loco-manip-161-105-textop.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "TextOp 主要解决数据闭环：用语言指令、遥操作/外骨骼数据采集人类操作和机器人状态，再通过扩散策略/流匹配、全身控制器/WBC/MPC转成可训练、可复用的全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-106-n106.md b/wiki/entities/paper-loco-manip-161-106-n106.md index c2ad94190..9cf241091 100644 --- a/wiki/entities/paper-loco-manip-161-106-n106.md +++ b/wiki/entities/paper-loco-manip-161-106-n106.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从语言指令、相机图像/多视角观测、人类视频/动捕轨迹恢复场景、目标或运动表征，再用扩散策略/流匹配、MM-DiT/Transformer 动作头、IK/动作重定向生成全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-107-n107.md b/wiki/entities/paper-loco-manip-161-107-n107.md index 2556dfaaa..b7781d87e 100644 --- a/wiki/entities/paper-loco-manip-161-107-n107.md +++ b/wiki/entities/paper-loco-manip-161-107-n107.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、本体状态与关节序列、仿真交互数据恢复场景、目标或运动表征，再用扩散策略/流匹配、VLM 语义规划/路由生成全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-108-n108.md b/wiki/entities/paper-loco-manip-161-108-n108.md index 2d009e28b..870b8a758 100644 --- a/wiki/entities/paper-loco-manip-161-108-n108.md +++ b/wiki/entities/paper-loco-manip-161-108-n108.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从人类视频/动捕轨迹恢复场景、目标或运动表征，再用扩散策略/流匹配生成全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-109-falcon.md b/wiki/entities/paper-loco-manip-161-109-falcon.md index a3bd7c688..820ed500e 100644 --- a/wiki/entities/paper-loco-manip-161-109-falcon.md +++ b/wiki/entities/paper-loco-manip-161-109-falcon.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "FALCON 主要解决数据闭环：用本体状态与关节序列、人类视频/动捕轨迹、遥操作/外骨骼数据采集人类操作和机器人状态，再通过全身控制器/WBC/MPC转成可训练、可复用的关节位置/力矩命令、末端执行器/腕手目标。关键点是把全身控制器/WBC/MPC放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-110-hdmi.md b/wiki/entities/paper-loco-manip-161-110-hdmi.md index 6430554fa..b7ed1e8d7 100644 --- a/wiki/entities/paper-loco-manip-161-110-hdmi.md +++ b/wiki/entities/paper-loco-manip-161-110-hdmi.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HDMI 先从相机图像/多视角观测、本体状态与关节序列、人类视频/动捕轨迹恢复场景、目标或运动表征，再用IK/动作重定向、分层技能/专家策略生成全身轨迹/动作序列。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-111-hitter.md b/wiki/entities/paper-loco-manip-161-111-hitter.md index f88900e7c..cca792ff9 100644 --- a/wiki/entities/paper-loco-manip-161-111-hitter.md +++ b/wiki/entities/paper-loco-manip-161-111-hitter.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HITTER 先从人类视频/动捕轨迹、接触力/触觉信号恢复场景、目标或运动表征，再用PPO/RL 策略训练、ACT/行为克隆模仿学习、全身控制器/WBC/MPC生成全身轨迹/动作序列、低层控制器目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-112-humanx.md b/wiki/entities/paper-loco-manip-161-112-humanx.md index 437fd977c..7d97770c9 100644 --- a/wiki/entities/paper-loco-manip-161-112-humanx.md +++ b/wiki/entities/paper-loco-manip-161-112-humanx.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HumanX 先从人类视频/动捕轨迹恢复场景、目标或运动表征，再用ACT/行为克隆模仿学习、IK/动作重定向、分层技能/专家策略生成可执行动作命令。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-113-humanoid.md b/wiki/entities/paper-loco-manip-161-113-humanoid.md index 3928790d6..b3cbece61 100644 --- a/wiki/entities/paper-loco-manip-161-113-humanoid.md +++ b/wiki/entities/paper-loco-manip-161-113-humanoid.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Humanoid 主要解决数据闭环：用人类视频/动捕轨迹、遥操作/外骨骼数据、接触力/触觉信号采集人类操作和机器人状态，再通过PPO/RL 策略训练、AMP/运动先验、扩散策略/流匹配转成可训练、可复用的可执行动作命令。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-115-resmimic.md b/wiki/entities/paper-loco-manip-161-115-resmimic.md index 9109d8227..0f19b2fc2 100644 --- a/wiki/entities/paper-loco-manip-161-115-resmimic.md +++ b/wiki/entities/paper-loco-manip-161-115-resmimic.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "ResMimic 先从本体状态与关节序列、人类视频/动捕轨迹、仿真交互数据恢复场景、目标或运动表征，再用策略网络和控制模块生成全身轨迹/动作序列。关键点是把策略网络和控制模块放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-116-wococo.md b/wiki/entities/paper-loco-manip-161-116-wococo.md index f1ca15590..ff554ff6c 100644 --- a/wiki/entities/paper-loco-manip-161-116-wococo.md +++ b/wiki/entities/paper-loco-manip-161-116-wococo.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "WoCoCo 先从本体状态与关节序列、人类视频/动捕轨迹、仿真交互数据恢复场景、目标或运动表征，再用PPO/RL 策略训练、AMP/运动先验、扩散策略/流匹配生成全身轨迹/动作序列、末端执行器/腕手目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-117-n117.md b/wiki/entities/paper-loco-manip-161-117-n117.md index 427bb456f..f946e38eb 100644 --- a/wiki/entities/paper-loco-manip-161-117-n117.md +++ b/wiki/entities/paper-loco-manip-161-117-n117.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、本体状态与关节序列、人类视频/动捕轨迹恢复场景、目标或运动表征，再用PPO/RL 策略训练、AMP/运动先验、全身控制器/WBC/MPC生成全身轨迹/动作序列、低层控制器目标。关键点是把PPO/RL 策略训练、AMP/运动先验放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-118-n118.md b/wiki/entities/paper-loco-manip-161-118-n118.md index fa60d626d..a18337d6d 100644 --- a/wiki/entities/paper-loco-manip-161-118-n118.md +++ b/wiki/entities/paper-loco-manip-161-118-n118.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从本体状态与关节序列、人类视频/动捕轨迹、仿真交互数据恢复场景、目标或运动表征，再用扩散策略/流匹配生成全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-119-n119.md b/wiki/entities/paper-loco-manip-161-119-n119.md index 1b17c0bc6..bff94cd79 100644 --- a/wiki/entities/paper-loco-manip-161-119-n119.md +++ b/wiki/entities/paper-loco-manip-161-119-n119.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、本体状态与关节序列、人类视频/动捕轨迹恢复场景、目标或运动表征，再用PPO/RL 策略训练生成关节位置/力矩命令、地形/场景表征。关键点是把PPO/RL 策略训练放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-120-gentlehumanoid.md b/wiki/entities/paper-loco-manip-161-120-gentlehumanoid.md index 52d363ca8..8efecc4fd 100644 --- a/wiki/entities/paper-loco-manip-161-120-gentlehumanoid.md +++ b/wiki/entities/paper-loco-manip-161-120-gentlehumanoid.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "GentleHumanoid 先从相机图像/多视角观测、本体状态与关节序列、人类视频/动捕轨迹恢复场景、目标或运动表征，再用PPO/RL 策略训练、全身控制器/WBC/MPC生成末端执行器/腕手目标。关键点是把PPO/RL 策略训练、全身控制器/WBC/MPC放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-121-humanoid-vla.md b/wiki/entities/paper-loco-manip-161-121-humanoid-vla.md index 69c6d0233..46b45369b 100644 --- a/wiki/entities/paper-loco-manip-161-121-humanoid-vla.md +++ b/wiki/entities/paper-loco-manip-161-121-humanoid-vla.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Humanoid-VLA 的实现路径是先把相机图像/多视角观测、人类视频/动捕轨迹编码成多模态表征，再用扩散策略/流匹配、VLA 多模态动作模型、全身控制器/WBC/MPC预测全身轨迹/动作序列。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-122-leverb.md b/wiki/entities/paper-loco-manip-161-122-leverb.md index c94b2b33b..a726c8cc3 100644 --- a/wiki/entities/paper-loco-manip-161-122-leverb.md +++ b/wiki/entities/paper-loco-manip-161-122-leverb.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "LeVERB 的实现路径是先把语言指令、相机图像/多视角观测、仿真交互数据编码成多模态表征，再用ACT/行为克隆模仿学习、VLA 多模态动作模型、潜变量/动作 token预测全身轨迹/动作序列、动作 chunk/token、低层控制器目标。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-123-n123.md b/wiki/entities/paper-loco-manip-161-123-n123.md index 7d120ede9..e9f3f6034 100644 --- a/wiki/entities/paper-loco-manip-161-123-n123.md +++ b/wiki/entities/paper-loco-manip-161-123-n123.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、本体状态与关节序列、仿真交互数据恢复场景、目标或运动表征，再用教师-学生知识迁移、PPO/RL 策略训练、全身控制器/WBC/MPC生成可执行动作命令。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-124-n124.md b/wiki/entities/paper-loco-manip-161-124-n124.md index bedaac765..44a5baf8d 100644 --- a/wiki/entities/paper-loco-manip-161-124-n124.md +++ b/wiki/entities/paper-loco-manip-161-124-n124.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、人类视频/动捕轨迹、仿真交互数据恢复场景、目标或运动表征，再用扩散策略/流匹配、IK/动作重定向、分层技能/专家策略生成可执行动作命令。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-125-n125.md b/wiki/entities/paper-loco-manip-161-125-n125.md index 35b306718..265c50b10 100644 --- a/wiki/entities/paper-loco-manip-161-125-n125.md +++ b/wiki/entities/paper-loco-manip-161-125-n125.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、本体状态与关节序列、仿真交互数据恢复场景、目标或运动表征，再用教师-学生知识迁移、PPO/RL 策略训练、全身控制器/WBC/MPC生成可执行动作命令。关键点是用特权信息训练教师策略，再把能力蒸馏到只能使用部署观测的学生策略。" related: diff --git a/wiki/entities/paper-loco-manip-161-126-n126.md b/wiki/entities/paper-loco-manip-161-126-n126.md index 0011a33fa..fe98ad5d6 100644 --- a/wiki/entities/paper-loco-manip-161-126-n126.md +++ b/wiki/entities/paper-loco-manip-161-126-n126.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、本体状态与关节序列、人类视频/动捕轨迹恢复场景、目标或运动表征，再用异构动捕与合成平衡数据、全身控制器/WBC/MPC、分层技能/专家策略生成低层控制器目标、导航/到达目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-127-n127.md b/wiki/entities/paper-loco-manip-161-127-n127.md index 797767dc0..98db188a1 100644 --- a/wiki/entities/paper-loco-manip-161-127-n127.md +++ b/wiki/entities/paper-loco-manip-161-127-n127.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、本体状态与关节序列、深度/点云/高度图恢复场景、目标或运动表征，再用PPO/RL 策略训练、扩散策略/流匹配、世界模型/视频预测生成低层控制器目标、地形/场景表征。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-128-bifrostumi.md b/wiki/entities/paper-loco-manip-161-128-bifrostumi.md index fba81b20d..37c3b4106 100644 --- a/wiki/entities/paper-loco-manip-161-128-bifrostumi.md +++ b/wiki/entities/paper-loco-manip-161-128-bifrostumi.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "BifrostUMI 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、扩散策略/流匹配、DINO/视觉特征抽取转成可训练、可复用的全身轨迹/动作序列、低层控制器目标、地形/场景表征。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-129-homie.md b/wiki/entities/paper-loco-manip-161-129-homie.md index a0c5523d1..57566247c 100644 --- a/wiki/entities/paper-loco-manip-161-129-homie.md +++ b/wiki/entities/paper-loco-manip-161-129-homie.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HOMIE 主要解决数据闭环：用本体状态与关节序列、遥操作/外骨骼数据、仿真交互数据采集人类操作和机器人状态，再通过PPO/RL 策略训练、全身控制器/WBC/MPC转成可训练、可复用的地形/场景表征。关键点是把PPO/RL 策略训练、全身控制器/WBC/MPC放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-130-humanoidexo.md b/wiki/entities/paper-loco-manip-161-130-humanoidexo.md index d5f82aa96..22a99ce5f 100644 --- a/wiki/entities/paper-loco-manip-161-130-humanoidexo.md +++ b/wiki/entities/paper-loco-manip-161-130-humanoidexo.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "HumanoidExo 主要解决数据闭环：用人类视频/动捕轨迹、遥操作/外骨骼数据采集人类操作和机器人状态，再通过扩散策略/流匹配、全身控制器/WBC/MPC、分层技能/专家策略转成可训练、可复用的全身轨迹/动作序列。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-131-open-television.md b/wiki/entities/paper-loco-manip-161-131-open-television.md index 97b51d472..11a2a2a5a 100644 --- a/wiki/entities/paper-loco-manip-161-131-open-television.md +++ b/wiki/entities/paper-loco-manip-161-131-open-television.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Open-TeleVision 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、扩散策略/流匹配、MM-DiT/Transformer 动作头转成可训练、可复用的关节位置/力矩命令、末端执行器/腕手目标、动作 chunk/token。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执" related: diff --git a/wiki/entities/paper-loco-manip-161-132-twist2.md b/wiki/entities/paper-loco-manip-161-132-twist2.md index e23f8e647..76eea6f8c 100644 --- a/wiki/entities/paper-loco-manip-161-132-twist2.md +++ b/wiki/entities/paper-loco-manip-161-132-twist2.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "TWIST2 的实现路径是先把相机图像/多视角观测、人类视频/动捕轨迹、遥操作/外骨骼数据编码成多模态表征，再用ACT/行为克隆模仿学习、VLA 多模态动作模型、全身控制器/WBC/MPC预测可执行动作命令。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-133-twist.md b/wiki/entities/paper-loco-manip-161-133-twist.md index 5c099f947..18111e40a 100644 --- a/wiki/entities/paper-loco-manip-161-133-twist.md +++ b/wiki/entities/paper-loco-manip-161-133-twist.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "TWIST 主要解决数据闭环：用人类视频/动捕轨迹、遥操作/外骨骼数据、接触力/触觉信号采集人类操作和机器人状态，再通过异构动捕与合成平衡数据、PPO/RL 策略训练、ACT/行为克隆模仿学习转成可训练、可复用的全身轨迹/动作序列、低层控制器目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-134-wt-umi.md b/wiki/entities/paper-loco-manip-161-134-wt-umi.md index 978a9d281..e1b5814f6 100644 --- a/wiki/entities/paper-loco-manip-161-134-wt-umi.md +++ b/wiki/entities/paper-loco-manip-161-134-wt-umi.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "WT-UMI 主要解决数据闭环：用相机图像/多视角观测、遥操作/外骨骼数据、接触力/触觉信号采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、扩散策略/流匹配转成可训练、可复用的全身轨迹/动作序列、末端执行器/腕手目标、动作 chunk/token。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-135-amo.md b/wiki/entities/paper-loco-manip-161-135-amo.md index d5d95f191..33bf59557 100644 --- a/wiki/entities/paper-loco-manip-161-135-amo.md +++ b/wiki/entities/paper-loco-manip-161-135-amo.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "AMO 把本体状态与关节序列、仿真交互数据、接触力/触觉信号转成可跟踪的身体目标，并通过PPO/RL 策略训练、ACT/行为克隆模仿学习、全身控制器/WBC/MPC训练或组合全身策略，最终输出全身轨迹/动作序列。关键点是把示范轨迹压成可监督的动作预测问题，再通过动作 chunk 或闭环执行降低时序抖动。" related: diff --git a/wiki/entities/paper-loco-manip-161-136-demohlm.md b/wiki/entities/paper-loco-manip-161-136-demohlm.md index 742177e4c..b5d4d48ae 100644 --- a/wiki/entities/paper-loco-manip-161-136-demohlm.md +++ b/wiki/entities/paper-loco-manip-161-136-demohlm.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "DemoHLM 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、全身控制器/WBC/MPC、闭环纠错/人类干预转成可训练、可复用的关节位置/力矩命令、全身轨迹/动作序列、低层控制器目标。关键点是把示范轨迹压成可监督的动作预测问题，再通过动作 chunk 或闭环执行降低时序抖动。" related: diff --git a/wiki/entities/paper-loco-manip-161-137-gallant.md b/wiki/entities/paper-loco-manip-161-137-gallant.md index 24191afa3..96c7ab358 100644 --- a/wiki/entities/paper-loco-manip-161-137-gallant.md +++ b/wiki/entities/paper-loco-manip-161-137-gallant.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Gallant 先从相机图像/多视角观测、本体状态与关节序列、深度/点云/高度图恢复场景、目标或运动表征，再用策略网络和控制模块生成地形/场景表征、导航/到达目标。关键点是把策略网络和控制模块放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-138-mobile-television.md b/wiki/entities/paper-loco-manip-161-138-mobile-television.md index 1442cb559..2bb189d0c 100644 --- a/wiki/entities/paper-loco-manip-161-138-mobile-television.md +++ b/wiki/entities/paper-loco-manip-161-138-mobile-television.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Mobile-TeleVision 把本体状态与关节序列、人类视频/动捕轨迹、接触力/触觉信号转成可跟踪的身体目标，并通过PPO/RL 策略训练、AMP/运动先验、扩散策略/流匹配训练或组合全身策略，最终输出关节位置/力矩命令、全身轨迹/动作序列、地形/场景表征。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-139-open-television.md b/wiki/entities/paper-loco-manip-161-139-open-television.md index fc189bac8..27099dea8 100644 --- a/wiki/entities/paper-loco-manip-161-139-open-television.md +++ b/wiki/entities/paper-loco-manip-161-139-open-television.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Open-TeleVision 主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、扩散策略/流匹配、MM-DiT/Transformer 动作头转成可训练、可复用的关节位置/力矩命令、末端执行器/腕手目标、动作 chunk/token。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执" related: diff --git a/wiki/entities/paper-loco-manip-161-140-twist2.md b/wiki/entities/paper-loco-manip-161-140-twist2.md index 50b1c7fe7..3809f43aa 100644 --- a/wiki/entities/paper-loco-manip-161-140-twist2.md +++ b/wiki/entities/paper-loco-manip-161-140-twist2.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "TWIST2 的实现路径是先把相机图像/多视角观测、人类视频/动捕轨迹、遥操作/外骨骼数据编码成多模态表征，再用ACT/行为克隆模仿学习、VLA 多模态动作模型、全身控制器/WBC/MPC预测可执行动作命令。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-141-toddlerbot.md b/wiki/entities/paper-loco-manip-161-141-toddlerbot.md index 9c0a8473c..551b8fde5 100644 --- a/wiki/entities/paper-loco-manip-161-141-toddlerbot.md +++ b/wiki/entities/paper-loco-manip-161-141-toddlerbot.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "ToddlerBot 主要解决数据闭环：用本体状态与关节序列、遥操作/外骨骼数据、仿真交互数据采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、扩散策略/流匹配、分层技能/专家策略转成可训练、可复用的地形/场景表征。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-142-n142.md b/wiki/entities/paper-loco-manip-161-142-n142.md index 62826164b..e88100560 100644 --- a/wiki/entities/paper-loco-manip-161-142-n142.md +++ b/wiki/entities/paper-loco-manip-161-142-n142.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、本体状态与关节序列、深度/点云/高度图恢复场景、目标或运动表征，再用下视深度相机和 U-Net 高度图重建、步态相位/频率调节、教师-学生知识迁移生成关节位置/力矩命令、地形/场景表征。关键点是把地形重建、步态相位和全身姿态放进同一个控制回路，而不是把感知和运控拆成松散串联。" related: diff --git a/wiki/entities/paper-loco-manip-161-143-act.md b/wiki/entities/paper-loco-manip-161-143-act.md index bf44dc3e2..15e65a6d6 100644 --- a/wiki/entities/paper-loco-manip-161-143-act.md +++ b/wiki/entities/paper-loco-manip-161-143-act.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "在人形操作的ACT模仿学习中，主动立体相机的性能优于多传感器设置 把相机图像/多视角观测、本体状态与关节序列、接触力/触觉信号转成可跟踪的身体目标，并通过ACT/行为克隆模仿学习、MM-DiT/Transformer 动作头训练或组合全身策略，最终输出动作 chunk/token。关键点是把示范轨迹压成可监督的动作预测问题，再通过动作 chunk 或闭环执行降低时序抖动。" related: diff --git a/wiki/entities/paper-loco-manip-161-144-n144.md b/wiki/entities/paper-loco-manip-161-144-n144.md index 6fa549434..0575d1a36 100644 --- a/wiki/entities/paper-loco-manip-161-144-n144.md +++ b/wiki/entities/paper-loco-manip-161-144-n144.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作主要解决数据闭环：用相机图像/多视角观测、本体状态与关节序列、遥操作/外骨骼数据采集人类操作和机器人状态，再通过ACT/行为克隆模仿学习、扩散策略/流匹配、MM-DiT/Transformer 动作头转成可训练、可复用的关节位置/力矩命令、全身轨迹/动作序列、末端执行器/腕手目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-145-n145.md b/wiki/entities/paper-loco-manip-161-145-n145.md index 4e01c1374..1d28366f8 100644 --- a/wiki/entities/paper-loco-manip-161-145-n145.md +++ b/wiki/entities/paper-loco-manip-161-145-n145.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先从相机图像/多视角观测、本体状态与关节序列、人类视频/动捕轨迹恢复场景、目标或运动表征，再用异构动捕与合成平衡数据、全身控制器/WBC/MPC、分层技能/专家策略生成低层控制器目标、地形/场景表征、导航/到达目标。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-146-agibot-world-colosseo.md b/wiki/entities/paper-loco-manip-161-146-agibot-world-colosseo.md index 54eb4a0c3..6d7d27fdc 100644 --- a/wiki/entities/paper-loco-manip-161-146-agibot-world-colosseo.md +++ b/wiki/entities/paper-loco-manip-161-146-agibot-world-colosseo.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "AgiBot World Colosseo 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用VLM 语义规划/路由、潜变量/动作 token、MM-DiT/Transformer 动作头预测全身轨迹/动作序列、末端执行器/腕手目标、动作 chunk/token。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-147-dial.md b/wiki/entities/paper-loco-manip-161-147-dial.md index bb617d4aa..39b3923dd 100644 --- a/wiki/entities/paper-loco-manip-161-147-dial.md +++ b/wiki/entities/paper-loco-manip-161-147-dial.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "DIAL 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用扩散策略/流匹配、VLA 多模态动作模型、VLM 语义规划/路由预测全身轨迹/动作序列、动作 chunk/token、地形/场景表征。关键点是让视频/世界模型提供可预测的物理先验，再由动作头把语义目标变成连续控制。" related: diff --git a/wiki/entities/paper-loco-manip-161-148-gr00t-n1.md b/wiki/entities/paper-loco-manip-161-148-gr00t-n1.md index ed873aec9..823b387a6 100644 --- a/wiki/entities/paper-loco-manip-161-148-gr00t-n1.md +++ b/wiki/entities/paper-loco-manip-161-148-gr00t-n1.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "GR00T N1 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用ACT/行为克隆模仿学习、扩散策略/流匹配、VLA 多模态动作模型预测可执行动作命令。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-149-gemini-robotics.md b/wiki/entities/paper-loco-manip-161-149-gemini-robotics.md index b3be69fb4..53db6075e 100644 --- a/wiki/entities/paper-loco-manip-161-149-gemini-robotics.md +++ b/wiki/entities/paper-loco-manip-161-149-gemini-robotics.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Gemini Robotics 的实现路径是先把语言指令、相机图像/多视角观测编码成多模态表征，再用ACT/行为克隆模仿学习、扩散策略/流匹配、VLA 多模态动作模型预测全身轨迹/动作序列、动作 chunk/token。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-150-genie-envisioner.md b/wiki/entities/paper-loco-manip-161-150-genie-envisioner.md index 4cdb3a367..f71bb3a36 100644 --- a/wiki/entities/paper-loco-manip-161-150-genie-envisioner.md +++ b/wiki/entities/paper-loco-manip-161-150-genie-envisioner.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Genie Envisioner 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用ACT/行为克隆模仿学习、扩散策略/流匹配、世界模型/视频预测预测地形/场景表征。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-151-legs.md b/wiki/entities/paper-loco-manip-161-151-legs.md index 79398fae0..27445c332 100644 --- a/wiki/entities/paper-loco-manip-161-151-legs.md +++ b/wiki/entities/paper-loco-manip-161-151-legs.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "LEGS 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用ACT/行为克隆模仿学习、VLA 多模态动作模型、世界模型/视频预测预测低层控制器目标、地形/场景表征。关键点是让视频/世界模型提供可预测的物理先验，再由动作头把语义目标变成连续控制。" related: diff --git a/wiki/entities/paper-loco-manip-161-152-lingbot-vla.md b/wiki/entities/paper-loco-manip-161-152-lingbot-vla.md index b5127f927..dd0d83750 100644 --- a/wiki/entities/paper-loco-manip-161-152-lingbot-vla.md +++ b/wiki/entities/paper-loco-manip-161-152-lingbot-vla.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "LingBot-VLA 的实现路径是先把语言指令、相机图像/多视角观测编码成多模态表征，再用VLA 多模态动作模型、VLM 语义规划/路由预测可执行动作命令。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-153-metaworld-x.md b/wiki/entities/paper-loco-manip-161-153-metaworld-x.md index b22adb623..40ce87d6d 100644 --- a/wiki/entities/paper-loco-manip-161-153-metaworld-x.md +++ b/wiki/entities/paper-loco-manip-161-153-metaworld-x.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "MetaWorld-X 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用PPO/RL 策略训练、AMP/运动先验、VLM 语义规划/路由预测地形/场景表征。关键点是把任务拆成可路由的技能或专家策略，再用高层模块在执行中选择和组合。" related: diff --git a/wiki/entities/paper-loco-manip-161-154-openhlm.md b/wiki/entities/paper-loco-manip-161-154-openhlm.md index eac607b4b..5ed2ae6a5 100644 --- a/wiki/entities/paper-loco-manip-161-154-openhlm.md +++ b/wiki/entities/paper-loco-manip-161-154-openhlm.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "OpenHLM 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用ACT/行为克隆模仿学习、扩散策略/流匹配、VLA 多模态动作模型预测全身轨迹/动作序列、低层控制器目标、地形/场景表征。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-155-physiflow.md b/wiki/entities/paper-loco-manip-161-155-physiflow.md index 114d48e56..09e5bc65d 100644 --- a/wiki/entities/paper-loco-manip-161-155-physiflow.md +++ b/wiki/entities/paper-loco-manip-161-155-physiflow.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "PhysiFlow 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用扩散策略/流匹配、VLA 多模态动作模型、VLM 语义规划/路由预测全身轨迹/动作序列、动作 chunk/token。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-156-psi0.md b/wiki/entities/paper-loco-manip-161-156-psi0.md index 9527e9f52..36e382593 100644 --- a/wiki/entities/paper-loco-manip-161-156-psi0.md +++ b/wiki/entities/paper-loco-manip-161-156-psi0.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "Psi0 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用异构动捕与合成平衡数据、VLM 语义规划/路由、潜变量/动作 token预测动作 chunk/token。关键点是把异构动捕与合成平衡数据、VLM 语义规划/路由放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-157-refine-dp.md b/wiki/entities/paper-loco-manip-161-157-refine-dp.md index 66cf6cd99..055632b66 100644 --- a/wiki/entities/paper-loco-manip-161-157-refine-dp.md +++ b/wiki/entities/paper-loco-manip-161-157-refine-dp.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "REFINE-DP 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用PPO/RL 策略训练、ACT/行为克隆模仿学习、扩散策略/流匹配预测关节位置/力矩命令、低层控制器目标。关键点是把动作生成看成条件生成问题，用扩散或流匹配在多模态动作分布里采样可执行轨迹。" related: diff --git a/wiki/entities/paper-loco-manip-161-158-rove.md b/wiki/entities/paper-loco-manip-161-158-rove.md index feadb444a..7e68dfc85 100644 --- a/wiki/entities/paper-loco-manip-161-158-rove.md +++ b/wiki/entities/paper-loco-manip-161-158-rove.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "ROVE 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用PPO/RL 策略训练、扩散策略/流匹配、VLA 多模态动作模型预测可执行动作命令。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-loco-manip-161-159-n159.md b/wiki/entities/paper-loco-manip-161-159-n159.md index 71f2921b5..dc1bf3068 100644 --- a/wiki/entities/paper-loco-manip-161-159-n159.md +++ b/wiki/entities/paper-loco-manip-161-159-n159.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用ACT/行为克隆模仿学习、扩散策略/流匹配、VLA 多模态动作模型预测动作 chunk/token。关键点是让视频/世界模型提供可预测的物理先验，再由动作头把语义目标变成连续控制。" related: diff --git a/wiki/entities/paper-loco-manip-161-160-n160.md b/wiki/entities/paper-loco-manip-161-160-n160.md index 749483bb1..bae85f8e4 100644 --- a/wiki/entities/paper-loco-manip-161-160-n160.md +++ b/wiki/entities/paper-loco-manip-161-160-n160.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "这篇工作先把语言指令、相机图像/多视角观测编码成多模态表征，再用策略网络和控制模块预测全身轨迹/动作序列。关键点是把策略网络和控制模块放在同一条训练/部署链路里，减少高层目标到低层动作之间的断点。" related: diff --git a/wiki/entities/paper-loco-manip-161-161-egovla.md b/wiki/entities/paper-loco-manip-161-161-egovla.md index 2c02f7937..1de54073c 100644 --- a/wiki/entities/paper-loco-manip-161-161-egovla.md +++ b/wiki/entities/paper-loco-manip-161-161-egovla.md @@ -2,7 +2,7 @@ type: entity tags: [paper, loco-manipulation, loco-manip-161-survey, humanoid] status: complete -updated: 2026-06-26 +updated: 2026-06-29 venue: curated summary: "EgoVLA 的实现路径是先把语言指令、相机图像/多视角观测、本体状态与关节序列编码成多模态表征，再用ACT/行为克隆模仿学习、VLA 多模态动作模型、潜变量/动作 token预测全身轨迹/动作序列、末端执行器/腕手目标。关键点是保留 VLM 的语义理解，同时增加机器人状态和动作头，避免只停留在语言规划。" related: diff --git a/wiki/entities/paper-sonic.md b/wiki/entities/paper-sonic.md index d30ee0b3c..ec032fa83 100644 --- a/wiki/entities/paper-sonic.md +++ b/wiki/entities/paper-sonic.md @@ -1,13 +1,16 @@ --- type: entity -tags: [paper, humanoid, rl, motion-control, body-system-stack, bfm, behavior-foundation-model, nvidia] +tags: [paper, humanoid, rl, motion-control, body-system-stack, bfm, behavior-foundation-model, nvidia, loco-manipulation, loco-manip-161-survey] status: complete -updated: 2026-06-26 +updated: 2026-06-29 arxiv: "2511.07820" venue: "2025 · arXiv" summary: "SONIC：规模化运动跟踪人形全身控制；在 RL 身体系统栈属参考跟踪层，在 BFM 谱系强调 goal-conditioned 与运控基座覆盖面。" related: + - ../overview/humanoid-loco-manip-161-papers-technology-map.md + - ../overview/loco-manip-161-category-01-motion-base-wbt.md + - ../overview/loco-manip-161-category-04-generative-language-trajectory.md - ../overview/humanoid-motion-cerebellum-technology-map.md - ../overview/motion-cerebellum-category-04-wbt-base.md - ../overview/humanoid-rl-motion-control-body-system-stack.md @@ -15,6 +18,7 @@ related: - ../overview/bfm-41-papers-technology-map.md - ../overview/bfm-category-02-goal-conditioned-learning.md - ../concepts/behavior-foundation-model.md + - ../tasks/loco-manipulation.md - ../methods/sonic-motion-tracking.md sources: - ../../sources/papers/humanoid_rl_stack_17_sonic_supersizing_motion_tracking_for_natural_hu.md @@ -25,6 +29,10 @@ sources: - ../../sources/blogs/wechat_embodied_ai_lab_bfm_41_papers_survey.md - ../../sources/papers/motion_cerebellum_64_catalog.md - ../../sources/blogs/wechat_embodied_ai_lab_humanoid_motion_cerebellum_survey.md + - ../../sources/papers/loco_manip_161_survey_019_sonic.md + - ../../sources/papers/loco_manip_161_survey_103_sonic.md + - ../../sources/blogs/wechat_embodied_ai_lab_humanoid_loco_manip_161_survey.md + - ../../sources/papers/humanoid_loco_manip_161_catalog.md --- # SONIC @@ -76,6 +84,17 @@ SONIC 的题目是 Supersizing Motion Tracking for Natural Humanoid Whole-Body C | 分组 | 02 Goal-conditioned 学习 | | 索引来源 | [awesome-bfm-papers](https://github.com/friedrichyuan/awesome-bfm-papers) | +### 在人形 Loco-Manip 161 篇中 + +同一篇论文在 [Loco-Manip 161 篇技术地图](../overview/humanoid-loco-manip-161-papers-technology-map.md) 里出现 **两次**（策展分类不同，canonical 实体仅此页）： + +| 槽位 | 分组 | 分类 hub | +|------|------|----------| +| 019/161 | 01 运控基座与通用全身跟踪 | [loco-manip-161-category-01-motion-base-wbt](../overview/loco-manip-161-category-01-motion-base-wbt.md) | +| 103/161 | 04 生成式运动、语言控制与轨迹规划 | [loco-manip-161-category-04-generative-language-trajectory](../overview/loco-manip-161-category-04-generative-language-trajectory.md) | + +索引来源：[具身智能研究室 · 161 篇人形 Loco-Manip 长文](https://mp.weixin.qq.com/s/pACh9EhsISiyPGdiiR0C3A) + ## 核心机制（归纳） ### 1）策展导读要点 @@ -108,6 +127,7 @@ SONIC 的方向很重要，但也要谨慎：运动控制的 scaling 不会和 - 方法深读：[sonic-motion-tracking.md](../methods/sonic-motion-tracking.md) - RL 身体系统栈：[humanoid-rl-motion-control-body-system-stack.md](../overview/humanoid-rl-motion-control-body-system-stack.md) - BFM 技术地图：[bfm-41-papers-technology-map.md](../overview/bfm-41-papers-technology-map.md) +- Loco-Manip 161 篇：[humanoid-loco-manip-161-papers-technology-map.md](../overview/humanoid-loco-manip-161-papers-technology-map.md) - BFM 概念：[behavior-foundation-model.md](../concepts/behavior-foundation-model.md) ## 参考来源 @@ -118,6 +138,10 @@ SONIC 的方向很重要，但也要谨慎：运动控制的 scaling 不会和 - [bfm_awesome_41_catalog.md](../../sources/papers/bfm_awesome_41_catalog.md) — 41+10 总表 - [wechat_embodied_ai_lab_humanoid_rl_motion_survey.md](../../sources/blogs/wechat_embodied_ai_lab_humanoid_rl_motion_survey.md) — RL 运动控制微信公众号编译导读 - [wechat_embodied_ai_lab_bfm_41_papers_survey.md](../../sources/blogs/wechat_embodied_ai_lab_bfm_41_papers_survey.md) — BFM 41 篇微信公众号编译导读 +- [loco_manip_161_survey_019_sonic.md](../../sources/papers/loco_manip_161_survey_019_sonic.md) — Loco-Manip 161 #019 策展摘录 +- [loco_manip_161_survey_103_sonic.md](../../sources/papers/loco_manip_161_survey_103_sonic.md) — Loco-Manip 161 #103 策展摘录 +- [humanoid_loco_manip_161_catalog.md](../../sources/papers/humanoid_loco_manip_161_catalog.md) — Loco-Manip 161 总表 +- [wechat_embodied_ai_lab_humanoid_loco_manip_161_survey.md](../../sources/blogs/wechat_embodied_ai_lab_humanoid_loco_manip_161_survey.md) — Loco-Manip 161 微信公众号编译导读 - 原始抓取：[wechat_humanoid_rl_42_survey_2026-05-26.md](../../sources/raw/wechat_humanoid_rl_42_survey_2026-05-26.md) - 论文： diff --git a/wiki/overview/humanoid-loco-manip-161-papers-technology-map.md b/wiki/overview/humanoid-loco-manip-161-papers-technology-map.md index bbb16b77f..cbc633973 100644 --- a/wiki/overview/humanoid-loco-manip-161-papers-technology-map.md +++ b/wiki/overview/humanoid-loco-manip-161-papers-technology-map.md @@ -2,7 +2,7 @@ type: overview tags: [loco-manipulation, humanoid, survey, embodied-ai, whole-body-control, vla] status: complete -updated: 2026-06-26 +updated: 2026-06-29 related: - ./loco-manip-161-category-01-motion-base-wbt.md - ./loco-manip-161-category-02-upper-body-interface.md diff --git a/wiki/overview/loco-manip-161-category-01-motion-base-wbt.md b/wiki/overview/loco-manip-161-category-01-motion-base-wbt.md index 4a59a7dbf..44d7a4864 100644 --- a/wiki/overview/loco-manip-161-category-01-motion-base-wbt.md +++ b/wiki/overview/loco-manip-161-category-01-motion-base-wbt.md @@ -2,7 +2,7 @@ type: overview tags: [loco-manipulation, humanoid, category-hub, survey] status: complete -updated: 2026-06-26 +updated: 2026-06-29 summary: "人形 Loco-Manip 161 篇 · 01 运控基座与通用全身跟踪（31 篇）— 底层身体控制、运动跟踪、抗扰动与通用动作执行。" related: - ./humanoid-loco-manip-161-papers-technology-map.md @@ -51,7 +51,7 @@ sources: | 016 | OmniH2O | [paper-loco-manip-161-016-omnih2o](../entities/paper-loco-manip-161-016-omnih2o.md) | | 017 | OmniRetarget | [paper-loco-manip-161-017-omniretarget](../entities/paper-loco-manip-161-017-omniretarget.md) | | 018 | Retargeting | [paper-loco-manip-161-018-retargeting](../entities/paper-loco-manip-161-018-retargeting.md) | -| 019 | SONIC | [paper-loco-manip-161-019-sonic](../entities/paper-loco-manip-161-019-sonic.md) | +| 019 | SONIC | [paper-sonic](../entities/paper-sonic.md) | | 020 | TWIST2 | [paper-loco-manip-161-020-twist2](../entities/paper-loco-manip-161-020-twist2.md) | | 021 | TWIST | [paper-loco-manip-161-021-twist](../entities/paper-loco-manip-161-021-twist.md) | | 022 | TextOp | [paper-loco-manip-161-022-textop](../entities/paper-loco-manip-161-022-textop.md) | diff --git a/wiki/overview/loco-manip-161-category-04-generative-language-trajectory.md b/wiki/overview/loco-manip-161-category-04-generative-language-trajectory.md index e10db762d..b3bd630df 100644 --- a/wiki/overview/loco-manip-161-category-04-generative-language-trajectory.md +++ b/wiki/overview/loco-manip-161-category-04-generative-language-trajectory.md @@ -2,7 +2,7 @@ type: overview tags: [loco-manipulation, humanoid, category-hub, survey] status: complete -updated: 2026-06-26 +updated: 2026-06-29 summary: "人形 Loco-Manip 161 篇 · 04 生成式运动、语言控制与轨迹规划（16 篇）— 从语言、目标或条件输入生成全身动作和轨迹。" related: - ./humanoid-loco-manip-161-papers-technology-map.md @@ -43,7 +43,7 @@ sources: | 100 | MotionWAM | [paper-loco-manip-161-100-motionwam](../entities/paper-loco-manip-161-100-motionwam.md) | | 101 | OMG | [paper-loco-manip-161-101-omg](../entities/paper-loco-manip-161-101-omg.md) | | 102 | Robot Motion Diffusion Model | [paper-loco-manip-161-102-robot-motion-diffusion-model](../entities/paper-loco-manip-161-102-robot-motion-diffusion-model.md) | -| 103 | SONIC | [paper-loco-manip-161-103-sonic](../entities/paper-loco-manip-161-103-sonic.md) | +| 103 | SONIC | [paper-sonic](../entities/paper-sonic.md) | | 104 | SafeFlow | [paper-loco-manip-161-104-safeflow](../entities/paper-loco-manip-161-104-safeflow.md) | | 105 | TextOp | [paper-loco-manip-161-105-textop](../entities/paper-loco-manip-161-105-textop.md) | | 106 | 从语言到运动 | [paper-loco-manip-161-106-n106](../entities/paper-loco-manip-161-106-n106.md) |