Model Card for "Decoder Only Transformer (DOT) Policy" for ALOHA cube transfer problem

Read more about the model and implementation details in the DOT Policy repository.

This model is trained using the LeRobot library and achieves state-of-the-art results on behavior cloning on ALOHA bimanual insert dataset. It achieves 92.6% success rate vs. 83% for the previous state-of-the-art model (ACT). (Note: it looks like the LeRobot implementation is not deterministic of environment makes it easier than the original problem, I am comparing it with https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human).

You can use this model by installing LeRobot from this branch

To train the model:

python lerobot/scripts/train.py \
    --policy.type=dot \
    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
    --env.type=aloha \
    --env.task=AlohaTransferCube-v0 \
    --output_dir=outputs/train/pusht_aloha_transfer_cube \
    --batch_size=24  \
    --log_freq=1000 \
    --eval_freq=5000 \
    --save_freq=5000 \
    --offline.steps=100000 \
    --seed=100000 \
    --wandb.enable=true \
    --num_workers=24 \
    --use_amp=true \
    --device=cuda \
    --policy.optimizer_lr=0.0001 \
    --policy.optimizer_min_lr=0.0001 \
    --policy.optimizer_lr_cycle_steps=100000 \
    --policy.train_horizon=75 \
    --policy.inference_horizon=50 \
    --policy.lookback_obs_steps=20 \
    --policy.lookback_aug=5 \
    --policy.rescale_shape="[480,640]" \
    --policy.alpha=0.98 \
    --policy.train_alpha=0.99 \
    --wandb.project=transfer_cube

To evaluate the model:

python lerobot/scripts/eval.py \
    --policy.path=jadechoghari/dot_transfer_cube \
    --env.type=aloha \
    --env.task=AlohaTransferCube-v0 \
    --eval.n_episodes=1000 \
    --eval.batch_size=100 \
    --seed=1000000

Model size:

• Total parameters: 14.1m
• Trainable parameters: 2.9m