LeHome Challenge — ACT Policy (act_pant_boosted, step 80K)
ACT policy for the LeHome bedroom garment-folding challenge.
Trained on the four_types_pant_boosted dataset, fine-tuned from act_four_types/100K.
Evaluation Results
Average success rate over 60 episodes per category (12 garments × 5 episodes).
| Category | Success Rate |
|---|---|
| top_long | 68.33% |
| top_short | 63.33% |
| pant_long | 41.67% |
| pant_short | 78.33% |
| Average | 62.92% |
Selected from candidate checkpoints (60K / 80K / 100K) based on:
- highest average success rate
- best balance across all 4 categories (highest minimum)
- avoidance of overfitting (top_long monotonically dropped after 80K)
Files
| File | Description |
|---|---|
policy.py |
ACT inference wrapper (BasePolicyServer subclass) |
server.py |
HTTP policy server (challenge protocol) |
Dockerfile |
CPU-only build (lerobot==0.4.3, torch CPU) |
pretrained_model/ |
Trained ACT policy (LeRobot v3 format) |
dataset/meta/ |
LeRobot dataset metadata (for LeRobotDatasetMetadata) |
constraints.txt |
Pinned versions |
Usage
docker build -t lehome-act-80k .
docker run --rm -p 8080:8080 lehome-act-80k
Then evaluate with the official challenge code:
python -m scripts.eval --policy_type docker \
--garment_type top_long --headless --device cpu --enable_cameras
Model Details
- Architecture: ACT (Action Chunking Transformer)
- Action / state dim: 12
- Image inputs: top_rgb, left_rgb, right_rgb (480×640×3)
- Training: 80K steps, batch size 16, fine-tuned from
act_four_types/100K - Dataset:
four_types_pant_boosted(1500 episodes, 397K frames, 4 garment categories with pant emphasis)