Skip to content

FF Master Cut Watermelon Workflow#504

Open
BOLT232 wants to merge 3 commits into
Faraday-Future-AI:mainfrom
BOLT232:main
Open

FF Master Cut Watermelon Workflow#504
BOLT232 wants to merge 3 commits into
Faraday-Future-AI:mainfrom
BOLT232:main

Conversation

@BOLT232

@BOLT232 BOLT232 commented Jun 27, 2026

Copy link
Copy Markdown

Registration UUID (required) · 参赛 UUID(必填)

Registration UUID: 63731519-ad75-4874-be78-71ec0e0bc957

Obtain your UUID from the official Robothon platform, and use the same UUID in submissions/<your-project>/registration.json.
从官方 Robothon 平台获取 UUID,并在 submissions/<你的项目>/registration.json 中填写同一个 UUID。

Project Summary · 项目说明

  • Project name · 项目名称: FF Master Cut Watermelon Workflow
  • Robot platform · 机器人平台: MuJoCo 3.x
  • Task goal · 任务目标: Use the FF Master humanoid robot to perform a complete chef-style watermelon workflow: grasp a knife with a five-finger hand, cut a whole watermelon in half, reposition, cut again into quarters, and serve one quarter to a plate with stable whole-body coordination and physically grounded fracture dynamics.
  • Technical approach · 技术方案: This project is a MuJoCo-based whole-body manipulation benchmark built around a 17-phase autonomous controller and a 15-DOF five-finger hand. The right arm executes a sensor-adaptive cutting sequence (APPROACH → ALIGN → CONTACT → SLICE → RETRACT, repeated for two cuts), while the hand uses real finger–handle contact physics, proximal tactile sensing, knife axial compliance, and a closed-loop grip servo. The watermelon uses an 8-layer predefined fracture path with Griffith-style per-layer energy accumulation, a continuous crack-front HUD, and cohesive-zone-style release logic; fracture release and slab separation are driven by accumulated cut energy and depth-scaled impulse rather than a simple kinematic threshold. The full episode ends with quarter serving to a garnished plate, with episode-level logging, offscreen rendering, and summary metrics.
  • Core features · 核心功能:
    • 15-DOF five-finger knife grasp with biomimetic rolling closure.
    • Real finger–handle contact physics with proximal grip-force sensing.
    • Closed-loop grip servo with knife axial compliance.
    • Sensor-adaptive dual-cut controller with 17 distinct phases.
    • Work-integral cut trigger and live blade/contact sensing.
    • 8-layer progressive fracture system with Griffith-style energy criterion.
    • Continuous crack-front HUD and cohesive-zone-style fracture release.
    • Whole-body coordination with waist lean, head gaze, and left-arm stabilization gesture.
    • Quarter serving workflow with weld-stabilized plating animation.
    • Offscreen demo video rendering, episode logging, and optional per-step CSV export.
  • Highlights · 亮点:
    • Complete kitchen-style workflow in one episode: whole watermelon → halves → quarters → plated serving.
    • Fracture is no longer a single instant release; it progresses through an 8-layer predefined crack path with per-layer energy accumulation.
    • The knife is held by a physically grounded five-finger grasp with tactile feedback and adaptive grip control.
    • ALIGN → CONTACT transitions are sensor-adaptive rather than purely time-based.
    • The submission is fully reproducible from the provided code, scene assets, and rendering script.
  • Current limitations · 当前局限:
    • The crack path is predefined; this is not free-topology FEM tearing.
    • The knife is still position-actuated rather than simulated as a fully deformable cutting tool.
    • Fracture is modeled as an energy-based layered rigid-body proxy, not continuous material failure.
  • Future improvements · 未来改进:
    • Extend from predefined fracture layers to more flexible crack-path evolution.
    • Add richer perception outputs, such as segmentation or depth data, to the data-collection pipeline.
    • Expand the kitchen workflow to additional food-prep tasks and stronger bimanual interaction.
    • Compare the current 15-DOF hand against alternative dexterous hand models.

How to Run · 如何运行

conda activate robothon

# Install dependencies
pip install -r requirements.txt

# Record demo video
python3 scripts/record_robot_video.py

# Optional: collect per-step CSVs and summary JSON
python3 scripts/record_robot_video.py --collect

# Optional: change rendering settings
python3 scripts/record_robot_video.py --fps 60 --width 1280 --height 720

Demo Video · 演示视频

  • Demo video is included in the submission folder · 演示视频已包含在提交文件夹中
  • Demo video link · 演示视频链接:

Checklist · 检查清单

  • submissions/<your-project>/registration.json contains my UUID · 已包含我的 UUID
  • PR description contains the same UUID · PR 描述中填写了同一个 UUID
  • Code runs from documented instructions · 代码可按文档说明运行
  • Demo video was generated by the submitted code · 演示视频由提交的代码生成

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant