diff --git a/README.md b/README.md index 8ceec49fa..044d8546d 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-1513节点_10183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_10195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 eff03b10c..c11b197ff 100644 --- a/docs/exports/graph-stats.json +++ b/docs/exports/graph-stats.json @@ -1 +1 @@ -{"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 +{"generated_at":"2026-06-29","node_count":1513,"edge_count":10195,"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":268,"人形论文深读笔记（Humanoid Paper Notebooks） 社区":252,"移动操作（Loco-Manipulation） 社区":236,"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":0.167,"community_quality_warning":false},"latest_wiki_nodes":[{"path":"wiki/entities/paper-notebook-vmp.md","detail_id":"entity-paper-notebook-vmp","label":"VMP（Versatile Motion Priors）","type":"entity","recency":"2026-06-29","source":"log.md"},{"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"}],"latest_wiki_node":{"path":"wiki/entities/paper-notebook-vmp.md","detail_id":"entity-paper-notebook-vmp","label":"VMP（Versatile Motion Priors）","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 87a7ebda9..59dad3253 100644 --- a/docs/exports/home-stats.json +++ b/docs/exports/home-stats.json @@ -1,13 +1,21 @@ { "generated_at": "2026-06-29", "node_count": 1513, - "edge_count": 10183, + "edge_count": 10195, "coverage": { "covered": 1482, "total": 1511, "percent": 98 }, "latest_wiki_nodes": [ + { + "path": "wiki/entities/paper-notebook-vmp.md", + "detail_id": "entity-paper-notebook-vmp", + "label": "VMP（Versatile Motion Priors）", + "type": "entity", + "recency": "2026-06-29", + "source": "log.md" + }, { "path": "wiki/entities/paper-sonic.md", "detail_id": "entity-paper-sonic", @@ -159,20 +167,12 @@ "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", + "path": "wiki/entities/paper-notebook-vmp.md", + "detail_id": "entity-paper-notebook-vmp", + "label": "VMP（Versatile Motion Priors）", "type": "entity", "recency": "2026-06-29", "source": "log.md" diff --git a/docs/index.html b/docs/index.html index 9bcd31c7f..c7481d6b8 100644 --- a/docs/index.html +++ b/docs/index.html @@ -43,7 +43,7 @@

系统学习人形机器人运动控制

1513 Nodes · - 10183 Links · + 10195 Links · 1482/1511 Sources
diff --git a/exports/graph-stats.json b/exports/graph-stats.json index eff03b10c..c11b197ff 100644 --- a/exports/graph-stats.json +++ b/exports/graph-stats.json @@ -1 +1 @@ -{"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 +{"generated_at":"2026-06-29","node_count":1513,"edge_count":10195,"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":268,"人形论文深读笔记（Humanoid Paper Notebooks） 社区":252,"移动操作（Loco-Manipulation） 社区":236,"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":0.167,"community_quality_warning":false},"latest_wiki_nodes":[{"path":"wiki/entities/paper-notebook-vmp.md","detail_id":"entity-paper-notebook-vmp","label":"VMP（Versatile Motion Priors）","type":"entity","recency":"2026-06-29","source":"log.md"},{"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"}],"latest_wiki_node":{"path":"wiki/entities/paper-notebook-vmp.md","detail_id":"entity-paper-notebook-vmp","label":"VMP（Versatile Motion Priors）","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 87a7ebda9..59dad3253 100644 --- a/exports/home-stats.json +++ b/exports/home-stats.json @@ -1,13 +1,21 @@ { "generated_at": "2026-06-29", "node_count": 1513, - "edge_count": 10183, + "edge_count": 10195, "coverage": { "covered": 1482, "total": 1511, "percent": 98 }, "latest_wiki_nodes": [ + { + "path": "wiki/entities/paper-notebook-vmp.md", + "detail_id": "entity-paper-notebook-vmp", + "label": "VMP（Versatile Motion Priors）", + "type": "entity", + "recency": "2026-06-29", + "source": "log.md" + }, { "path": "wiki/entities/paper-sonic.md", "detail_id": "entity-paper-sonic", @@ -159,20 +167,12 @@ "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", + "path": "wiki/entities/paper-notebook-vmp.md", + "detail_id": "entity-paper-notebook-vmp", + "label": "VMP（Versatile Motion Priors）", "type": "entity", "recency": "2026-06-29", "source": "log.md" diff --git a/log.md b/log.md index 38432ecac..8ad37ab97 100644 --- a/log.md +++ b/log.md @@ -1,5 +1,7 @@ > 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] ingest | sources/papers/humanoid_pnb_vmp.md — VMP β-VAE motion prior + 条件 PPO 全身跟踪；wiki/entities/paper-notebook-vmp.md；交叉 whole-body-tracking-pipeline / character-animation-vs-robotics / humanoid-motion-tracking-method-selection + ## [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) diff --git a/sources/papers/humanoid_pnb_vmp.md b/sources/papers/humanoid_pnb_vmp.md index 7d3ce4128..db0ce837a 100644 --- a/sources/papers/humanoid_pnb_vmp.md +++ b/sources/papers/humanoid_pnb_vmp.md @@ -1,27 +1,80 @@ # VMP: Versatile Motion Priors for Robustly Tracking Motion on Physical Characters -> 来源归档（ingest · Humanoid Paper Notebooks progress 待深读） +> 来源归档（ingest · Disney Research PDF） - **标题：** VMP: Versatile Motion Priors for Robustly Tracking Motion on Physical Characters -- **类型：** paper -- **深读状态：** 待撰写（见 [progress.json](https://github.com/ImChong/Humanoid_Robot_Learning_Paper_Notebooks/blob/main/progress.json)） -- **计划笔记路径：** `papers/04_Loco-Manipulation_and_WBC/VMP__Versatile_Motion_Priors_for_Robustly_Tracking_Motion_on_Physical_Characters/VMP__Versatile_Motion_Priors_for_Robustly_Tracking_Motion_on_Physical_Characters.md` -- **分类：** 04_Loco-Manipulation_and_WBC -- **路线：** Loco-Manipulation -- **入库日期：** 2026-06-11 -- **一句话说明：** 列入 Paper Notebooks 阅读进度，深读笔记尚未完成；本文件为 **进度 → wiki** 溯源锚点。 +- **类型：** paper / physics-based-animation / motion-tracking / motion-prior / character-control / sim2real +- **机构：** ETH Zürich、Disney Research（瑞士） +- **作者：** Agon Serifi, Ruben Grandia, Espen Knoop, Markus Gross, Moritz Bächer +- **发表：** SCA 2024（ACM SIGGRAPH/Eurographics Symposium on Computer Animation，Montreal，2024-08） +- **DOI：** +- **PDF：** +- **项目页：** +- **Paper Notebooks 分类：** 04_Loco-Manipulation_and_WBC（待深读清单锚点） +- **入库日期：** 2026-06-11（PNB 锚点）；2026-06-29（PDF 深读 ingest） +- **一句话说明：** 两阶段「β-VAE 运动潜空间 + 条件 PPO 跟踪」：从未过滤大规模动捕学 versatile motion prior，用全身运动学参考 + 时变 latent 驱动单一策略，在仿真与 LIME 双足真机上实现高精度、可艺术家导向的物理角色跟踪。 -## 核心摘录（策展，非全文） +## 相关资料（策展） -- 本文件锚定 **待深读** 论文在姊妹仓库 `progress.json` 中的条目；笔记完成后应改用笔记页链接并深化 wiki 归纳。 -- 知识归纳见 wiki 实体页：[paper-notebook-vmp](../../wiki/entities/paper-notebook-vmp.md). +| 类型 | 链接 | 说明 | +|------|------|------| +| PDF | | 本次 ingest 主来源 | +| DOI | | Computer Graphics Forum / SCA 2024 | +| ETH 页 | | 摘要与下载 | +| 对照 | CALM [TKG*23]、ASE [PGH*22]、DeepMimic [PALVdP18] | 端到端 latent vs 显式跟踪 vs 对抗先验 | +| 硬件线 | BDX / LIME 角色机器人 [GKH*24] | 同一 Disney 双足角色研究线 | + +## 摘要级要点 + +- **问题：** 从未结构化动捕学单一控制策略，使其对**多样且未见**运动保持高精度跟踪，并可部署到真机，仍很困难。 +- **两阶段解耦：** Stage I 用 **β-VAE** 在短运动窗口上自监督学 kinematic latent；Stage II 用 **PPO** 训练条件策略 $\pi(a_t|s_t, m_t, z_t)$，$m_t$ 为当前参考帧、$z_t$ 为以该帧为中心的窗口 latent——分离自监督表征与显式 imitation reward，避免对抗训练的 mode collapse。 +- **接口：** 用户/艺术家提供 **全身运动学参考序列**；encoder 映射为时变 latent 流，policy 输出力矩/PD 目标，实现 physics-informed 全身控制。 +- **规模：** **11 h** 未过滤数据（CMU 8.5 h + Mixamo 2 h + Reallusion 0.5 h）；36-DoF 人形 + **LIME** 20-DoF 双足（0.84 m，16.2 kg）真机验证。 +- **相对 CALM：** latent 可分性更高（LDA **0.854 vs 0.687**）；完整管线 RTX 4090 **<3 天** vs CALM A100 **~2 周**；未见舞蹈/格斗跟踪误差约 **5°** 关节 MAE（全数据训练）。 + +## 核心摘录（面向 wiki 编译） + +### 1) Stage I：Versatile Motion Prior（β-VAE） + +- **运动窗口：** $M_t=\{m_{t-W},\ldots,m_{t+W}\}$，$W=30$（约 1 s）；每帧 $m_t=\{h_t,\theta_t,v_t,q_t,\dot q_t,p_t\}$（根高、6D 朝向、速度、关节角/速、手足相对根位置）。 +- **归一化：** 以中心帧根朝向为局部 heading frame；除朝向外其余量按数据集均值/方差归一化。 +- **网络：** 1D Conv + ConvResNet blocks；$d_z=64$；β-VAE，$\beta=0.002$，KL 周期调度；RTX 4090 训练 **~10 h**。 +- **设计要点：** **每帧一个 latent**（非整段单码），便于细粒度控制、即时响应参考突变与空间组合。 + +### 2) Stage II：条件 PPO 跟踪策略 + +- **条件：** $c_t=(m_t, z_t)$；策略为 3 层 512 MLP + 对角高斯；**Isaac Gym** 8192 并行环境，PPO **~48 h**。 +- **奖励：** $r_t=r^{\text{track}}_t+r^{\text{alive}}_t+r^{\text{smooth}}_t$——根/关节/末端跟踪 MSE + 存活 + 动作一阶/二阶平滑与力矩惩罚。 +- **终止：** 末端偏差超阈持续 $f$ 帧则 early terminate（非仅足接触终止），允许全身着地。 +- **鲁棒性：** 质量/摩擦/推扰域随机化；**执行器模型**（PD 电机 + Coulomb 摩擦 + 速度相关力矩限）；机器人额外关节标定噪声 $\epsilon_q$。 + +### 3) 消融与对比（人形，Reallusion 子集 + 未见集） + +| 条件输入 | Idle | Walk | Attack | Dance | Unseen（关节 MAE °） | +|----------|------|------|--------|-------|----------------------| +| M（仅当前帧） | 7.52 | 8.63 | 13.11 | 12.79 | 13.29 | +| L（仅 latent） | 6.10 | 7.53 | 10.70 | 10.45 | 10.92 | +| **LM（本文）** | **4.31** | **4.15** | **7.08** | **5.80** | **7.83** | + +- 全数据训练后未见动作 MAE **~5°**；对不可行参考（空中爬楼梯）尽量跟踪并保持平衡。 +- 相对 CALM：训练更快、latent 信息更丰富；LM 耦合使策略对参考更「跟手」。 + +### 4) 艺术家导向与真机 + +- **空间组合：** 不同 clip 的臂/身可拼接；**运动编辑：** 任意排序片段后精调关键帧/风格化。 +- **LIME 真机：** 无踝 roll 时策略自适应用脚尖触地维持平衡；动态动作在物理执行器极限内仍高保真跟踪。 +- **局限：** 长规划视野特技（后空翻等）需更强记忆结构；生成式遍历 latent 空间尚未展开。 ## 对 wiki 的映射 - [paper-notebook-vmp](../../wiki/entities/paper-notebook-vmp.md) +- [Whole-Body Tracking Pipeline](../../wiki/concepts/whole-body-tracking-pipeline.md) +- [Character Animation vs Robotics](../../wiki/concepts/character-animation-vs-robotics.md) +- [DeepMimic（方法）](../../wiki/methods/deepmimic.md) +- [人形运动跟踪方法选型](../../wiki/queries/humanoid-motion-tracking-method-selection.md) - 分类父节点：[paper-notebook-category-04-loco-manipulation-and-wbc](../../wiki/overview/paper-notebook-category-04-loco-manipulation-and-wbc.md) ## 参考来源（原始） -- [Humanoid Robot Learning Paper Notebooks · progress.json](https://github.com/ImChong/Humanoid_Robot_Learning_Paper_Notebooks/blob/main/progress.json) - +- [VMP PDF](https://la.disneyresearch.com/wp-content/uploads/VMP_paper.pdf) — 本次 ingest 主来源 +- [Humanoid Robot Learning Paper Notebooks · progress.json](https://github.com/ImChong/Humanoid_Robot_Learning_Paper_Notebooks/blob/main/progress.json) — 待深读清单锚点 diff --git a/wiki/concepts/character-animation-vs-robotics.md b/wiki/concepts/character-animation-vs-robotics.md index 598537594..42ea627f7 100644 --- a/wiki/concepts/character-animation-vs-robotics.md +++ b/wiki/concepts/character-animation-vs-robotics.md @@ -3,7 +3,7 @@ type: concept tags: [humanoid, character-animation, entertainment-robotics, motion-retargeting, reward-design, style-prior] status: complete created: 2026-05-19 -updated: 2026-06-12 +updated: 2026-06-29 related: - ./motion-retargeting.md - ./motion-retargeting-pipeline.md @@ -12,6 +12,7 @@ related: - ../entities/paper-deeprl-locomotion-action-space-sca2017.md - ../queries/legged-humanoid-rl-pd-gain-setting.md - ../methods/disney-olaf-character-robot.md + - ../entities/paper-notebook-vmp.md - ../methods/deepmimic.md - ../methods/amp-reward.md - ../methods/ase.md @@ -25,6 +26,7 @@ sources: - ../../sources/sites/blender-org.md - ../../sources/repos/blender.md - ../../sources/papers/disney_olaf_character_robot.md + - ../../sources/papers/humanoid_pnb_vmp.md - ../../sources/papers/amp.md - ../../sources/papers/deepmimic.md - ../../sources/sites/botlab_motioncanvas.md @@ -190,7 +192,7 @@ flowchart LR - 上游链路与方法谱系：[Motion Retargeting](./motion-retargeting.md)、[Motion Retargeting Pipeline](./motion-retargeting-pipeline.md)、[GMR vs NMR vs ReActor](../comparisons/gmr-vs-nmr-vs-reactor.md)。 - 安全与奖励配比：[Reward Design](./reward-design.md)、[Control Barrier Function](./control-barrier-function.md)。 -- 实机案例：[Disney Olaf 角色机器人](../methods/disney-olaf-character-robot.md)；中性平台 [Roboto Origin](../entities/roboto-origin.md)、[Asimov v1](../entities/asimov-v1.md)。 +- 实机案例：[Disney Olaf 角色机器人](../methods/disney-olaf-character-robot.md)、[VMP](../entities/paper-notebook-vmp.md)（LIME 双足 + 动画参考接口）；中性平台 [Roboto Origin](../entities/roboto-origin.md)、[Asimov v1](../entities/asimov-v1.md)。 - 风格先验方法：[DeepMimic](../methods/deepmimic.md)、[AMP](../methods/amp-reward.md)、[ASE](../methods/ase.md)；原作者索引 [Xue Bin Peng](../entities/xue-bin-peng.md)。 - **室内人–场景交互合成（运动学角色）：** [DIMOS](../entities/paper-dimos-human-scene-motion-synthesis.md)（ICCV 2023）用 RL 在 CVAE 运动基元潜空间上合成走–坐–躺序列，服务 AR/VR 与训练数据规模化；上真机需经重定向与接触动力学重建模，与 DeepMimic/AMP 的「仿真物理角色 → 机器人」迁移链正交。 - 工具层：[BotLab / MotionCanvas](../entities/botlab-motioncanvas.md)、[机器人关键帧与运动编辑工具](../entities/robot-motion-keyframe-editors.md)。 @@ -204,6 +206,7 @@ flowchart LR ## 参考来源 - [sources/papers/disney_olaf_character_robot.md](../../sources/papers/disney_olaf_character_robot.md) — Olaf 实机角色（arXiv:2512.16705） +- [sources/papers/humanoid_pnb_vmp.md](../../sources/papers/humanoid_pnb_vmp.md) — VMP β-VAE motion prior + 条件跟踪（SCA 2024） - [sources/papers/deepmimic.md](../../sources/papers/deepmimic.md) — DeepMimic 显式跟踪奖励 - [sources/papers/amp.md](../../sources/papers/amp.md) — AMP 判别器风格先验 - [sources/sites/botlab_motioncanvas.md](../../sources/sites/botlab_motioncanvas.md) — BotLab / MotionCanvas 浏览器节点图 diff --git a/wiki/concepts/whole-body-tracking-pipeline.md b/wiki/concepts/whole-body-tracking-pipeline.md index c140403f5..9afe1e1fb 100644 --- a/wiki/concepts/whole-body-tracking-pipeline.md +++ b/wiki/concepts/whole-body-tracking-pipeline.md @@ -3,7 +3,7 @@ type: concept tags: [robotics, humanoid, whole-body-tracking, wbt, pipeline, motion-tracking, cross-embodiment, sim2real] status: complete created: 2026-05-29 -updated: 2026-06-25 +updated: 2026-06-29 summary: "Whole-Body Tracking（WBT）端到端流水线：参考采集 → 重定向 → 训练数据 → 策略学习 → 跨具身迁移 → 真机部署的统一视图，对比 SONIC / BeyondMimic / SD-AMP / Heracles / Any2Any / GMT(RGMT) 等 6 条主流落地路径在每一阶段的取舍。" related: - ./motion-retargeting-pipeline.md @@ -24,6 +24,7 @@ related: - ../entities/paper-perceptive-bfm.md - ../entities/paper-hrl-stack-14-robust_and_generalized_humanoid_moti.md - ../entities/paper-resmimic.md + - ../entities/paper-notebook-vmp.md - ../queries/humanoid-motion-tracking-method-selection.md - ../overview/humanoid-rl-motion-control-body-system-stack.md - ../entities/sam-3d-body.md @@ -38,6 +39,7 @@ sources: - ../../sources/papers/humanoid_rl_stack_17_sonic_supersizing_motion_tracking_for_natural_hu.md - ../../sources/papers/humanoid_rl_stack_15_beyondmimic_from_motion_tracking_to_versatile_hu.md - ../../sources/papers/humanoid_rl_stack_40_heracles_bridging_precise_tracking_and_generativ.md + - ../../sources/papers/humanoid_pnb_vmp.md --- # Whole-Body Tracking Pipeline（全身运动跟踪流水线） @@ -172,6 +174,7 @@ WBT 的核心分歧在**奖励/损失**怎么写。四条主流： | 路线 | 核心思想 | 代表 | |------|----------|------| | **显式 tracking 奖励** | 分项 reward 直接对齐参考（位姿/速度/末端） | [DeepMimic](../methods/deepmimic.md)、[BeyondMimic](../methods/beyondmimic.md) | +| **预训练 latent + 显式跟踪** | β-VAE 运动先验与 PPO 解耦，动画参考接口 | [VMP](../entities/paper-notebook-vmp.md) | | **对抗式 motion prior** | 判别器约束风格分布，tracking 由任务奖励驱动 | [AMP](../methods/amp-reward.md)、[SD-AMP](../entities/paper-unified-walk-run-recovery-sdamp.md) | | **生成式中间件** | 扩散/VAE 在 tracking 失败时改写参考 | [Heracles](../entities/paper-heracles-humanoid-diffusion.md) | | **规模化通用 tracker** | 海量参考 + 大模型，把 tracking 当**预训练任务** | [SONIC](../methods/sonic-motion-tracking.md)、[GMT/RGMT](../entities/paper-hrl-stack-14-robust_and_generalized_humanoid_moti.md)、[Any2Track](../methods/any2track.md) | @@ -204,8 +207,10 @@ WBT 的核心分歧在**奖励/损失**怎么写。四条主流： | **Heracles** | 跟踪失败时由扩散重生成 | — | tracking + 扩散中间件兜底 | 单具身 | 项目页 demo | [Heracles](../entities/paper-heracles-humanoid-diffusion.md) | | **Any2Any** | 复用源机参考池 | 运动学对齐层 | LoRA 后训练（约 1% 算力） | 多目标机 LoRA 适配 | LimX Oli/Luna、G1、H1 | [Any2Any](../entities/paper-any2any-cross-embodiment-wbt.md) | | **GMT (RGMT)** | 多任务参考 + 扰动课程 | 上游通用 | 历史编码 + 命令交叉注意力 | 单具身（强抗扰） | Unitree G1 项目页 | [RGMT](../entities/paper-hrl-stack-14-robust_and_generalized_humanoid_moti.md)、[Any2Track](../methods/any2track.md) | +| **VMP** | 11 h 未过滤 CMU/Mixamo/Reallusion | IK retarget（LIME） | β-VAE prior + 条件 PPO 跟踪 | 单具身角色平台 | LIME 双足真机 | [VMP](../entities/paper-notebook-vmp.md) | > **选型直觉**： +> - **动画工作流 + 角色真机** → VMP（kinematic 参考 + latent 上下文） > - **想"开箱即用通用 tracker"** → SONIC / GMT 系（高代价但泛化最好） > - **预算紧、想跑通一条线** → BeyondMimic 单具身重训（最成熟） > - **要少参考做多行为** → SD-AMP（3 条参考覆盖走/跑/起身） diff --git a/wiki/entities/paper-notebook-vmp.md b/wiki/entities/paper-notebook-vmp.md index 2ab33bc09..76f9fb766 100644 --- a/wiki/entities/paper-notebook-vmp.md +++ b/wiki/entities/paper-notebook-vmp.md @@ -1,62 +1,150 @@ --- type: entity -tags: [paper, humanoid-paper-notebooks, paper-notebook-planned] -status: planned -updated: 2026-06-26 -venue: "2024.08" +tags: [paper, humanoid-paper-notebooks, motion-tracking, motion-prior, physics-based-animation, reinforcement-learning, character-animation, eth, disney, sim2real] +status: complete +updated: 2026-06-29 +venue: "SCA 2024" related: - ../overview/paper-notebook-category-04-loco-manipulation-and-wbc.md - ../overview/humanoid-paper-notebooks-index.md + - ../concepts/whole-body-tracking-pipeline.md + - ../concepts/character-animation-vs-robotics.md + - ../methods/deepmimic.md + - ../methods/ase.md + - ../queries/humanoid-motion-tracking-method-selection.md + - ../entities/paper-loco-manip-161-102-robot-motion-diffusion-model.md sources: - ../../sources/papers/humanoid_pnb_vmp.md -summary: "VMP：列入 Paper Notebooks progress 待深读清单；深读笔记完成后升格为完整索引实体。" +summary: "VMP（SCA 2024）：β-VAE 从未过滤动捕学 versatile motion prior，再条件 PPO 跟踪全身运动学参考；单一策略覆盖 11 h 多样技能，LIME 双足真机验证动态跟踪与艺术家导向接口。" --- -# VMP +# VMP（Versatile Motion Priors） -**VMP: Versatile Motion Priors for Robustly Tracking Motion on Physical Characters** 已列入 [Humanoid Robot Learning Paper Notebooks](https://imchong.github.io/Humanoid_Robot_Learning_Paper_Notebooks/index.html) 的 **progress 待深读** 清单（分类：04_Loco-Manipulation_and_WBC）。本页为 **计划索引实体**，深读笔记尚未撰写；笔记完成后应链向笔记站并深化归纳。 +**VMP**（*Versatile Motion Priors for Robustly Tracking Motion on Physical Characters*，ETH Zürich · Disney Research，SCA 2024）提出 **两阶段解耦** 的物理角色控制：**Stage I** 用 **β-VAE** 在短运动窗口上自监督提取 kinematic latent；**Stage II** 用 **PPO** 学条件策略，输入为 **当前参考帧 + 时变 latent**，输出动力学一致的执行器指令。论文在 **11 小时未过滤** 人形动捕上训练 **单一策略**，支持未见用户指定运动、艺术家空间组合/编辑，并在 **LIME** 双足真机上演示高动态跟踪。 ## 一句话定义 -VMP 的人形机器人学习论文条目，当前处于 Paper Notebooks 阅读进度（待深读）阶段。 +VMP 用「**预训练运动潜空间 + 显式跟踪奖励**」替代端到端对抗 latent 或单 clip 专家，使一个 RL 策略在多样、未见运动学参考上保持高精度全身跟踪，并可直接对接动画工作流与真机双足平台。 ## 英文缩写速查 | 缩写 | 英文全称 | 简要说明 | |------|----------|----------| -| RL | Reinforcement Learning | 通过与环境交互最大化长期回报来学习策略 | -| WBC | Whole-Body Control | 协调全身关节满足多任务/约束的控制基础设施 | -| Sim2Real | Simulation to Real | 把仿真中学到的策略迁移落地真机的工程主线 | +| VMP | Versatile Motion Priors | 本文：从动捕学通用运动潜表征并条件下游跟踪 | +| VAE | Variational Autoencoder | Stage I：β-VAE 重建短窗口运动并产出 $z_t$ | +| PPO | Proximal Policy Optimization | Stage II：Isaac Gym 并行仿真中的 on-policy 训练 | +| RL | Reinforcement Learning | 第二阶段用跟踪+正则奖励学条件控制策略 | +| WBT | Whole-Body Tracking | 全身参考运动跟踪，与 DeepMimic 谱系同族 | +| Sim2Real | Simulation to Real | 域随机化+执行器模型后迁移 LIME 真机 | ## 为什么重要 -- 列入 Paper Notebooks **progress 待深读** 清单，便于与全库 [人形论文笔记总索引](../overview/humanoid-paper-notebooks-index.md) 及分类父节点交叉检索。 -- 在深读笔记完成前，本页作为 **占位子节点**，避免知识图谱缺失该论文实体。 +- **解耦表征与控制：** 相对 CALM/ASE 等 **端到端 latent+policy**，VMP 先用自监督 VAE 学 structured latent，再用 **显式 imitation reward** 训策略——论文报告更高 latent 可分性（LDA **0.854 vs 0.687**），并减轻对抗训练的 **mode collapse** 与长训成本（RTX 4090 **<3 天** vs CALM A100 **~2 周**）。 +- **单一策略 + 大库泛化：** 在 **CMU + Mixamo + Reallusion** 共 **11 h** **未过滤** 数据上，一个 MLP 策略跟踪 Idle/Walk/Attack/Dance 及未见序列；全数据训练后未见动作关节 MAE 约 **5°**。 +- **动画—机器人桥梁：** 控制接口是 **全身运动学参考序列**（非高层任务指令），支持 **空间组合、任意剪辑排序、风格化编辑**；与 Disney **BDX/LIME** 角色机器人线一致，证明图形学 motion prior 可落地物理双足。 +- **LM 条件缺一不可：** 消融显示仅当前帧（M）或仅 latent（L）均弱于 **$c_t=(m_t,z_t)$**；Dance 跟踪误差相对 M-only 约 **减半**。 + +## 流程总览 + +```mermaid +flowchart TB + subgraph s1 ["Stage I：Motion Prior（β-VAE）"] + clips["未过滤动捕库
CMU · Mixamo · Reallusion · 11 h"] + win["随机窗口 M_t
2W+1 帧，W=30"] + enc["Encoder e_ψ(z_t|M_t)"] + dec["Decoder 重建 M_t'"] + clips --> win --> enc --> dec + end + + subgraph s2 ["Stage II：条件 PPO"] + ref["用户/艺术家
运动学参考 m_t"] + zt["时变 latent z_t"] + pol["π(a_t | s_t, m_t, z_t)"] + sim["Isaac Gym 物理仿真"] + ref --> enc + enc --> zt + ref --> pol + zt --> pol + pol --> sim + end + + subgraph infer ["推理 / 真机"] + lime["LIME 双足 · 域随机化+执行器模型"] + sim --> lime + end +``` + +## 核心机制（归纳） + +### Stage I：短窗口 β-VAE + +| 要素 | 说明 | +|------|------| +| 状态 $m_t$ | 根高 $h_t$、6D 朝向 $\theta_t$、根/关节速度、关节角、手足相对根位置 $p_t$ | +| 窗口 | $M_t=\{m_{t-30},\ldots,m_{t+30}\}$，约 **1 s** 上下文 | +| 归一化 | 中心帧 heading 局部系 + 数据集均值方差（朝向除外） | +| 表征 | $d_z=64$；**每帧独立 latent 序列**（非整段单码），支持突变响应与空间组合 | +| 训练 | β=0.002，batch 512，RTX 4090 **~10 h** | + +### Stage II：条件跟踪策略 + +- **观测/条件：** 策略状态含本体传感 + 前两步动作；条件 $c_t=(m_t,z_t)$，$m_t$ 提供瞬时跟踪目标，$z_t$ 编码近过去/未来上下文。 +- **奖励：** 根位姿/速度、关节、末端跟踪 MSE + 存活项 + 动作平滑与力矩正则。 +- **终止：** 末端偏差持续超阈则终止（允许非足端着地），优于仅足接触终止的脆弱性。 +- **执行器：** PD 电机 + Coulomb 摩擦 + 速度相关力矩限；仿真与 LIME 使用 **Unitree-A1 / Dynamixel** 参数表。 +- **训练：** PPO，8192 并行 env，**~48 h** / RTX 4090。 + +### 艺术家导向接口 -## 核心信息 +- **空间组合：** 上身与下身可来自不同 clip（论文 Fig.6）。 +- **运动编辑：** 先拼接片段得初始序列，再精调关键事件位置/时机或风格化。 +- **交互推理：** encoder+policy 可在线运行，动画师可在非物理工具中编参考、即时得物理反馈（去脚滑等）。 -| 字段 | 内容 | +### 真机（LIME） + +- 20-DoF、0.84 m、16.2 kg；机载 IMU + 执行器编码器状态估计；动捕辅助标定。 +- 硬件缺踝 roll 时，策略用 **脚尖触地** 等自适应姿态逼近参考并保持平衡。 +- 动态踢腿、舞蹈类动作在物理力矩极限内仍保持风格跟踪。 + +## 实验与评测（索引级） + +| 设置 | 要点 | |------|------| -| 分类 | 04_Loco-Manipulation_and_WBC | -| 深读状态 | 待撰写（[progress.json](https://github.com/ImChong/Humanoid_Robot_Learning_Paper_Notebooks/blob/main/progress.json)） | -| 计划文件夹 | `papers/04_Loco-Manipulation_and_WBC/VMP__Versatile_Motion_Priors_for_Robustly_Tracking_Motion_on_Physical_Characters` | +| 角色 | 标准 **36-DoF** 人形；**LIME** 20-DoF 双足真机 | +| 数据 | Reallusion 0.5 h + CMU 8.5 h + Mixamo 2 h，**未过滤** | +| 消融 LM vs M/L | Dance 关节 MAE：**5.80°**（LM）vs **12.79°**（M）vs **10.45°**（L） | +| 泛化 | 全库训练后未见序列 **~5°** MAE；对不可行参考尽量跟踪且不倒 | +| vs CALM | 运动质量相当、artifact 更少；latent–参考耦合更强、重复动作更少 | +| 局限 | 长视野特技（后空翻）需带记忆结构；生成式 latent 遍历未演示 | +> 完整数值与视频以 [PDF](https://la.disneyresearch.com/wp-content/uploads/VMP_paper.pdf) 为准。 -## 实验与评测 +## 常见误区或局限 -- 深读笔记尚未完成；量化 benchmark、消融与实机指标待笔记撰写后补充。 +- **不是对抗 motion prior：** VMP 走 **显式跟踪 + 预训练 kinematic latent**，与 [AMP](../methods/amp-reward.md)/[ASE](../methods/ase.md) 的判别器路线目标不同——前者偏 **参考贴合**，后者偏 **风格分布**。 +- **≠ 通用 scaling tracker：** 相对 [SONIC](../methods/sonic-motion-tracking.md)/BeyondMimic 等人形 scaling 线，VMP 更强在 **动画工作流接口 + Disney 角色真机**，而非 AMASS 级人形工程栈。 +- **单 MLP 的记忆上限：** 论文承认需 hidden state 才能覆盖长飞行相特技；当前架构对 **即时跟踪** 足够，对 **长承诺特技** 不足。 +- **数据噪声未滤：** 刻意用未过滤 CMU 证明鲁棒性，但极端不可行动作仍会失败或近似跟踪。 ## 与其他页面的关系 +- [Whole-Body Tracking Pipeline](../concepts/whole-body-tracking-pipeline.md) — VMP 代表「**大库动捕 → latent prior → 显式跟踪 RL → 角色真机**」的动画侧 WBT 路径。 +- [DeepMimic](../methods/deepmimic.md) — 共享多 term 跟踪奖励与 RSI 式随机初始化；VMP 扩展为 **跨 clip 单一策略 + latent 条件**。 +- [Character Animation vs Robotics](../concepts/character-animation-vs-robotics.md) — 与 Disney Olaf/BDX 同属 **表演导向物理角色** 研究线。 +- [Robot Motion Diffusion Model](./paper-loco-manip-161-102-robot-motion-diffusion-model.md) — 同 Disney 系 **运动生成 + 物理角色** 互补（生成 vs 跟踪执行）。 +- [人形运动跟踪方法选型](../queries/humanoid-motion-tracking-method-selection.md) — 「需要 **latent 上下文 + 显式跟踪**、动画接口」时参考 VMP。 - 分类父节点：[paper-notebook-category-04-loco-manipulation-and-wbc](../overview/paper-notebook-category-04-loco-manipulation-and-wbc.md) -- 总索引：[humanoid-paper-notebooks-index.md](../overview/humanoid-paper-notebooks-index.md) ## 参考来源 -- [humanoid_pnb_vmp.md](../../sources/papers/humanoid_pnb_vmp.md) -- [Humanoid Robot Learning Paper Notebooks · progress.json](https://github.com/ImChong/Humanoid_Robot_Learning_Paper_Notebooks/blob/main/progress.json) - +- [humanoid_pnb_vmp.md](../../sources/papers/humanoid_pnb_vmp.md) — PDF 策展摘录（主来源） +- [VMP PDF](https://la.disneyresearch.com/wp-content/uploads/VMP_paper.pdf) +- [DOI 10.1111/cgf.15175](https://doi.org/10.1111/cgf.15175) +- [Humanoid Robot Learning Paper Notebooks · progress.json](https://github.com/ImChong/Humanoid_Robot_Learning_Paper_Notebooks/blob/main/progress.json) — 04_Loco-Manipulation 待深读锚点 ## 推荐继续阅读 -- [Paper Notebooks 阅读进度（PROGRESS.md）](https://github.com/ImChong/Humanoid_Robot_Learning_Paper_Notebooks/blob/main/papers/PROGRESS.md) +- [ETH CGL 论文页](https://cgl.ethz.ch/publications/papers/paperSer24a.php) +- [DeepMimic 方法页](../methods/deepmimic.md) — 显式跟踪奖励范式起点 +- [ASE](../methods/ase.md) — 端到端对抗 skill embedding 对照 +- [Disney Olaf 角色机器人](../methods/disney-olaf-character-robot.md) — 同机构娱乐型双足实机线 diff --git a/wiki/methods/deepmimic.md b/wiki/methods/deepmimic.md index f58e18b81..3e1825d34 100644 --- a/wiki/methods/deepmimic.md +++ b/wiki/methods/deepmimic.md @@ -2,7 +2,7 @@ type: method tags: [imitation-learning, tracking, rl, xbpeng] status: complete -updated: 2026-06-16 +updated: 2026-06-29 related: - ../entities/protomotions.md - ./amp-reward.md diff --git a/wiki/overview/paper-notebook-category-04-loco-manipulation-and-wbc.md b/wiki/overview/paper-notebook-category-04-loco-manipulation-and-wbc.md index 973bee762..052ba5e8f 100644 --- a/wiki/overview/paper-notebook-category-04-loco-manipulation-and-wbc.md +++ b/wiki/overview/paper-notebook-category-04-loco-manipulation-and-wbc.md @@ -2,7 +2,7 @@ type: overview tags: [humanoid-paper-notebooks, paper-index, overview] status: complete -updated: 2026-06-26 +updated: 2026-06-29 related: - ./humanoid-paper-notebooks-index.md summary: "Paper Notebooks 分类 04：运动操作与全身控制（158 篇深读笔记索引）。" @@ -171,7 +171,7 @@ summary: "Paper Notebooks 分类 04：运动操作与全身控制（158 篇深 - [VIGOR](../entities/paper-notebook-vigor-visual-goal-in-context-inference-for-unifi.md) — [深读笔记](https://imchong.github.io/Humanoid_Robot_Learning_Paper_Notebooks/papers/04_Loco-Manipulation_and_WBC/VIGOR_Visual_Goal-In-Context_Inference_for_Unified_Humanoid_Fall_Safety/VIGOR_Visual_Goal-In-Context_Inference_for_Unified_Humanoid_Fall_Safety.html) - [VIRAL](../entities/paper-hrl-stack-28-viral.md) — 待深读 - [VIRAL](../entities/paper-viral-humanoid-visual-sim2real.md) — 待深读 -- [VMP](../entities/paper-notebook-vmp.md) — 待深读 +- [VMP](../entities/paper-notebook-vmp.md) — β-VAE motion prior + 条件 PPO 全身跟踪（SCA 2024；LIME 真机） - [VisualMimic](../entities/paper-notebook-visualmimic.md) — 已 ingest（arXiv:2509.20322；视觉分层 sim2real + 关键点 tracker；深读笔记待撰写） - [Whole-Body Dynamic Throwing with Legged Manipulators](../entities/paper-notebook-whole-body-dynamic-throwing-with-legged-manipula.md) — 待深读 - [Whole-Body Model-Predictive Control of Legged Robots with MuJoCo](../entities/paper-notebook-whole-body-model-predictive-control-of-legged-ro.md) — 待深读 diff --git a/wiki/queries/humanoid-motion-tracking-method-selection.md b/wiki/queries/humanoid-motion-tracking-method-selection.md index a792bbee4..82ad259f9 100644 --- a/wiki/queries/humanoid-motion-tracking-method-selection.md +++ b/wiki/queries/humanoid-motion-tracking-method-selection.md @@ -7,6 +7,7 @@ updated: 2026-06-29 summary: 在人形 RL 运动控制栈中，如何按任务阶段在 DeepMimic / BeyondMimic / AMP 家族 / 通用 tracker / 接触丰富场景 tracking / 生成式动作先验之间选型。 sources: - ../../sources/papers/scenebot_arxiv_2606_27581.md + - ../../sources/papers/humanoid_pnb_vmp.md - ../../sources/papers/deepmimic.md - ../../sources/papers/amp.md - ../../sources/papers/smp.md @@ -52,6 +53,7 @@ flowchart TD | 证明「能跟参考跑起来」 | 显式 tracking reward | [DeepMimic](../methods/deepmimic.md)、[BeyondMimic](../methods/beyondmimic.md) | | 任务完成后仍像「人」 | 对抗式 motion prior | [AMP](../methods/amp-reward.md)、[ADD](../methods/add.md)、[SMP](../methods/smp.md) | | 多动作通用 tracker | 规模化 tracking policy | [Any2Track](../methods/any2track.md)、[AMS](../methods/ams.md)、[MotionBricks](../methods/motionbricks.md)、[EGM](../methods/egm-efficient-general-mimic.md)、[SONIC](../methods/sonic-motion-tracking.md)、[Humanoid-GPT](../entities/paper-humanoid-gpt.md) | +| 动画参考 + latent 上下文跟踪 | 两阶段 VAE prior + 显式 PPO | [VMP](../entities/paper-notebook-vmp.md)（SCA 2024；LIME 真机） | | 接触丰富场景 tracking | 参考运动 + per-link contact label | [SceneBot](../entities/paper-scenebot.md)（hindsight 场景重建 + 单策略 terrain/object） | | 数据稀缺、要合成参考 | 生成式动作 | [ASE](../methods/ase.md)、[GenMo](../methods/genmo.md)、[扩散动作生成](../methods/diffusion-motion-generation.md) | @@ -69,7 +71,7 @@ flowchart TD 当任务奖励已满足，仍出现步态不自然时，引入 [AMP](../methods/amp-reward.md) 判别器先验。[ADD](../methods/add.md) 用对抗差分减轻多目标手调；[SMP](../methods/smp.md) 走 **冻结扩散 + SDS** 路线（非判别器），先验预训练后可**丢弃原始 MoCap**、在多任务多策略间复用，代价是两阶段训练、同采样量 wall-clock 约为 AMP 的 ~1.8×（论文报告 600M samples：SMP ~11.5h vs AMP ~6.2h）。 -**选型轴**：每任务都要重训先验 / 必须保留数据集 → AMP/ADD；先验一次训好跨任务复用、不愿在 RL 阶段保留 MoCap → SMP。三者对比见 [AMP / ADD / SMP 运动先验变体对比](../comparisons/amp-add-smp-motion-prior-variants.md)。 +**选型轴**：每任务都要重训先验 / 必须保留数据集 → AMP/ADD；先验一次训好跨任务复用、不愿在 RL 阶段保留 MoCap → SMP；需要 **动画师可直接编 kinematic 参考**、且希望 **latent 与跟踪解耦**（相对 CALM/ASE 端到端）→ [VMP](../entities/paper-notebook-vmp.md)。三者对比见 [AMP / ADD / SMP 运动先验变体对比](../comparisons/amp-add-smp-motion-prior-variants.md)。 ### 3. 通用 tracker 与实时原语 @@ -149,6 +151,8 @@ flowchart TD ## 参考来源 +- [SceneBot（arXiv:2606.27581）](../../sources/papers/scenebot_arxiv_2606_27581.md) +- [VMP（SCA 2024 PDF）](../../sources/papers/humanoid_pnb_vmp.md) - [DeepMimic 论文摘要](../../sources/papers/deepmimic.md) - [AMP 论文摘要](../../sources/papers/amp.md) - [具身智能研究室：人形 AMP 先验综述](../../sources/blogs/wechat_embodied_ai_lab_humanoid_amp_motion_prior_survey.md) @@ -169,6 +173,7 @@ flowchart TD - [Heracles](../entities/paper-heracles-humanoid-diffusion.md)、[PhyGile](../entities/paper-phygile.md)、[SD-AMP](../entities/paper-unified-walk-run-recovery-sdamp.md)、[SPRINT](../entities/paper-sprint-humanoid-athletic-sprints.md) - [Any2Any](../entities/paper-any2any-cross-embodiment-wbt.md) - [SceneBot](../entities/paper-scenebot.md) +- [VMP](../entities/paper-notebook-vmp.md) ## 一句话记忆