Add script to convert LeRobot v3 datasets to unifolm-wma format

[brownwa]

Summary

• The HuggingFace datasets in the unitreerobotics collection (e.g. G1_WBT_Brainco_Pickup_Pillow) are published in LeRobot v3.0 format, but WMAData (in src/unifolm_wma/data/wma_data.py) expects a different layout — causing a FileNotFoundError on {dataset_name}.csv at inference time.
• Adds scripts/prepare_lerobot_dataset.py — a one-time conversion utility that reads the LeRobot parquet/video/stats files and produces the layout WMAData requires.
• The converted layout matches what DataModuleFromConfig.setup() constructs in src/unifolm_wma/utils/data.py.

What the script generates

{dataset_dir}/
  {dataset_name}.csv                                     # one row per episode
  videos/{dataset_name}/{episode_idx}.mp4                # per-episode clips (ffmpeg trim)
  transitions/{dataset_name}/{episode_idx}.h5            # state + action arrays
  transitions/{dataset_name}/meta_data/stats.safetensors # normalizer stats

Usage

python scripts/prepare_lerobot_dataset.py \
    --dataset_dir ~/.cache/huggingface/hub/datasets--unitreerobotics--G1_WBT_Brainco_Pickup_Pillow/snapshots/<hash> \
    --dataset_name G1_WBT_Brainco_Pickup_Pillow

Optional flags: --camera_key, --action_key, --state_key, --instruction.

Test plan

• Run prepare_lerobot_dataset.py on G1_WBT_Brainco_Pickup_Pillow
• Confirm WMAData loads 300 samples without error
• Confirm real_eval_server.py starts and serves successfully
• Test on at least one other dataset from the unitreerobotics collection

Related

Fixes the FileNotFoundError reported in issue #XX.

🤖 Generated with Claude Code

[brownwa]

All changes from this PR are now included in PR #47 (consolidated branch unifolm-wbt-dataset), which combines this PR with #45 and the Brainco setup guide into a single clean branch. Closing this in favour of #47.

[brownwa]

Superseded by PR #47.