lerobot-train \
    --policy.type=act \
    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human_image \
    --env.type=aloha \
    --env.task=AlohaTransferCube-v0 \
    --batch_size=32 \
    --steps=60000 \
    --eval_freq=5000 \
    --output_dir=./outputs/act_aloha_cube_best \
    --wandb.enable=false \
    --policy.push_to_hub=false

Evaluation

bash

lerobot-eval \
    --policy.path=LeTau/act_aloha_transfer_cube \
    --env.type=aloha \
    --env.task=AlohaTransferCube-v0 \
    --eval.batch_size=1 \
    --eval.n_episodes=20

Fine-tuning

bash

lerobot-train \
    --resume=true \
    --config_path=LeTau/act_aloha_transfer_cube/train_config.json \
    --steps=100000

Results

Evaluation	Episodes	Success Rate	Avg Sum Reward
Training	50	42%	116.26
Independent	20	35%	95.95

Expected success rate: 35-42%

Detailed Evaluation Results (Training)

Sum Rewards: [0.0, 241.0, 57.0, 201.0, 48.0, 0.0, 0.0, 220.0, 262.0, 0.0, 
              59.0, 211.0, 287.0, 187.0, 74.0, 2.0, 203.0, 18.0, 10.0, 0.0, 
              0.0, 263.0, 7.0, 57.0, 39.0, 214.0, 297.0, 24.0, 0.0, 274.0, 
              201.0, 2.0, 228.0, 228.0, 68.0, 290.0, 2.0, 222.0, 31.0, 219.0, 
              69.0, 22.0, 0.0, 76.0, 244.0, 227.0, 0.0, 26.0, 192.0, 211.0]

Successes: 21/50 episodes

Limitations

Limited training data: Only 50 demonstration episodes available
Moderate success rate: This is a lightweight baseline model
Single task: Only trained on TransferCube, no multi-task capability

Citation

bibtex

@article{zhao2023learning,
  title={Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware},
  author={Zhao, Tony Z and Kumar, Vikash and Levine, Sergey and Finn, Chelsea},
  journal={arXiv preprint arXiv:2304.13705},
  year={2023}
}

Acknowledgments

LeRobot framework by HuggingFace
ALOHA project by Stanford