ruali-dev/lehome-smolvla-submit-4types

LeHome Submission - SmolVLA (4 Types)

Method

We fine-tune lerobot/smolvla_base on the official LeHome merged four-garment dataset (four_types_merged).

We use a custom evaluation wrapper (custom_smolvla_policy.py) to remap LeHome RGB observation keys:

• observation.images.top_rgb -> observation.images.camera1
• observation.images.left_rgb -> observation.images.camera2
• observation.images.right_rgb -> observation.images.camera3

The policy takes:

• 12-dim robot state
• 3 RGB observations

and outputs:

• 12-dim bimanual joint actions.

Files

• pretrained_model/: trained checkpoint
• custom_smolvla_policy.py: custom evaluation wrapper
• rollout_results.txt: self-reported results

Environment

• Python 3.11
• Official LeHome repository
• LeRobot / SmolVLA
• Isaac Sim based evaluation pipeline

How to Evaluate

1. Clone the official LeHome repository.
2. Place custom_smolvla_policy.py under scripts/eval_policy/.
3. Import it in scripts/eval_policy/__init__.py.
4. Run evaluation:

python -m scripts.eval \
  --policy_type custom_smolvla \
  --policy_path <PATH_TO_PRETRAINED_MODEL> \
  --garment_type top_short \
  --dataset_root Datasets/example/four_types_merged \
  --num_episodes 1 \
  --enable_cameras \
  --device cpu \
  --headless

Replace top_short with:

• top_long
• pant_long
• pant_short

Notes

• This submission uses a custom wrapper because LeHome camera keys differ from the default SmolVLA expected keys.
• If datasets are removed, please keep the dataset metadata folders.