Dataset recording & format

Recording is covered in Teleoperation → Recording Demonstrations — this page covers what soarm_lerobot does with the result.

The data pipeline (dataset.py) — real, tested code

Function/class	Role
load_dataset(...)	Open a LeRobotDataset by repo-id or local root
compute_joint_stats(ds)	Per-joint mean/std arrays
JointNormalizer	Normalize/denormalize joint vectors from those stats
ChunkedActionDataset	A torch.utils.data.Dataset yielding action chunks (the ACT window; --chunk-size default 100)
create_dataloader(...)	Batching on top of the above

recorder.py's TeleopRecorder (the only symbol re-exported at package level) buffers per-frame joint, gripper, and task data from the teleop loop, detects episode boundaries from BTN_A toggling, and writes LeRobotDataset episodes as parquet + metadata — discarding short or noisy episodes automatically.

Package facts

• soarm_lerobot v0.1.0 · setuptools, flat layout · Python ≥3.10
• Hard dependencies: numpy, torch>=2.0, lerobot>=0.4, datasets
• Extras: [train] (diffusers, accelerate, wandb), [dev] (pytest, ruff)
• No tests yet — tests/ contains only __init__.py; the data pipeline is pure torch/numpy and hardware-free, so it's straightforward to cover.

See Training for the current state of the training loops that would consume this pipeline.