Skip to content

data_flow.md

Latest commit

 

History

History
25 lines (21 loc) · 1.11 KB

data_flow.md

File metadata and controls

25 lines (21 loc) · 1.11 KB

📈 Data Flow (TrackerLab)

The data flow in TrackerLab bridges human motion datasets with robotic trajectory execution and training via modular abstractions. Here's the typical pipeline:

graph LR
    A[AMASS / FBX Dataset] --> B[Preprocessing]
    B --> C[Retargeting<br>poselib/scripts/amass/retarget_all.py]
    C --> D[Retargeted Motions<br>data/retargeted/]
    D --> E[Trajectory Manager]
    E --> F[Skill FSM / Command Manager]
    F --> G[RL Env (e.g., H1TrackAll)]
    G --> H[Training<br>train.py]
Loading

Step-by-step:

  • Dataset (AMASS / FBX): Raw human motion sequences.
  • Preprocessing: T-pose normalization, joint filtering, frame truncation.
  • Retargeting: Convert motion to robot-specific format (using retarget_all.py).
  • Retargeted Output: Saved in data/retargeted/.
  • Trajectory Manager: Loads and manages motion playback with timing.
  • Skill FSM / Command Manager: Composes motions into tasks via state machines.
  • RL Environment: Registered as H1TrackAll, interfaces with IsaacLab simulation.
  • Training: Uses IsaacLab’s train.py to optimize policies.