Skip to content

data_collection.md

Latest commit

 

History

History
77 lines (57 loc) · 2.39 KB

data_collection.md

File metadata and controls

77 lines (57 loc) · 2.39 KB
Raw

Data collection

The collector runs the released RLTracker teacher in IsaacLab and writes the state-action trajectories consumed by PredActor training.

Inputs

Place the supplied motion set and teacher checkpoint at these default paths:

Artifacts/
├── checkpoints/rl_tracker/model_75000.pt
└── motions/g1_152fk/
    ├── motion_001/motion.npz
    └── ...

Download both inputs from the PredActor artifact repository:

uv run --locked python scripts/hf_download.py \
  --filter checkpoints/rl_tracker motions/g1_152fk

Install

Collection uses Python 3.10, Isaac Sim 4.5.0, and Isaac Lab 2.1.0 in a separate environment:

conda create --override-channels -c conda-forge -n predactor-collection python=3.10 -y
conda activate predactor-collection
bash PredActor/collection/install.sh

The installer keeps IsaacLab below PredActor/collection/.deps/ and installs the exported RLTracker task, RSL-RL runtime, and CUDA forward-kinematics library. It applies PredActor/collection/constraints.txt throughout installation and runs pip check before reporting success.

For an Aliyun Python package mirror, use:

PIP_INDEX_URL=https://mirrors.aliyun.com/pypi/simple \
  bash PredActor/collection/install.sh

PyTorch and Isaac Sim still use their CUDA/NVIDIA package indexes. After a network interruption, rerun the same installer; completed downloads and the IsaacLab checkout are reused.

Collect

Run deterministic collection from the public repository root. With the supplied 152 motions, 50 episodes per motion gives 7,600 episodes:

bash PredActor/collection/collect_dataset.sh \
  --mode deterministic \
  --samples_per_motion 50 \
  --num_envs 1000 \
  --headless

The output is Artifacts/dataset/g1_152fk_raw/motion.zarr. Command-line --checkpoint, --motion_file, and --output_dir values override those defaults.

For another motion set, use --motion_file "my_motions/*" under Artifacts/motions/. The episode target is the number of matched motions times --samples_per_motion. Use the same pattern for input.motion_pattern when dyeing the dataset.

Use --visualize for the native Isaac Sim viewer or --headless --livestream 2 for WebRTC. Run bash PredActor/collection/collect_dataset.sh --help for the complete collection and noise options.

Next, label the collected data, then train PredActor.