jadechoghari/dot_pusht_keypoints_best

Model Card for "Decoder Only Transformer (DOT) Policy" for PushT keypoints dataset

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 the PushT keypoints dataset. It achieves 94% success rate (and 0.985 average max reward) vs. ~78% for the previous state-of-the-art model or 69% that I managed to reproduce using VQ-BET implementation in LeRobot.

This is the best checkpoint for the model. These results are achievable assuming we have reliable validation and can select the best checkpoint based on the validation results (not always the case in robotics). If you are interested in more stable and reproducible results achievable without checkpoint selection, please refer to https://huggingface.co/IliaLarchenko/dotpushtkeypoints

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/pusht_keypoints \
    --env.type=pusht \
    --env.task=PushT-v0 \
    --output_dir=outputs/train/pusht_keyponts \
    --batch_size=24  \
    --log_freq=1000 \
    --eval_freq=10000 \
    --save_freq=50000 \
    --offline.steps=1000000 \
    --seed=100000 \
    --wandb.enable=true \
    --num_workers=24 \
    --use_amp=true \
    --device=cuda \
    --policy.return_every_n=2 \
    --policy.train_horizon=30 \
    --policy.inference_horizon=30

To evaluate the model:

python lerobot/scripts/eval.py \
    --policy.path=jadechoghari/dot_pusht_keypoints_best \
    --env.type=pusht \
    --env.task=PushT-v0 \
    --eval.n_episodes=1000 \
    --eval.batch_size=100 \
    --env.obs_type=environment_state_agent_pos \
    --seed=1000000

Model size:

• —Total parameters: 2.1m
• —Trainable parameters: 2.1m