Model Card for Diffusion Policy / PushT (keypoints)

Diffusion Policy (as per Diffusion Policy: Visuomotor Policy Learning via Action Diffusion) trained for the PushT environment from gym-pusht with keypoint-only observations.

Note: The original work trains keypoints-only with conditioning via inpainting. Here, we encode the observation along with the agent position and use the encoding as global conditioning for the denoising U-Net.

How to Get Started with the Model

Use python lerobot/scripts/eval.py -p lerobot/diffusion_pusht to evaluate for 50 episodes with the outputs sent to outputs/eval.

For further information, please see the LeRobot library (particularly the evaluation script).

Training Details

Trained with LeRobot@cc2f6e7.

The model was trained using [LeRobot's training script](TODO link with commit hash one the PR is merged) and with the pusht_keypoints dataset, using this command:

python lerobot/scripts/train.py \
  hydra.job.name=diffusion_pusht_keypoints \
  hydra.run.dir=outputs/train/2024-07-03/13-52-44_diffusion_pusht_keypoints \
  env=pusht_keypoints \
  policy=diffusion_pusht_keypoints \
  training.save_checkpoint=true \
  training.offline_steps=200000 \
  training.save_freq=20000 \
  training.eval_freq=10000 \
  training.log_freq=50 \
  training.num_workers=4 \
  eval.n_episodes=50 \
  eval.batch_size=50 \
  wandb.enable=true \
  wandb.disable_artifact=true \
  device=cuda \
  use_amp=true

The training curves may be found at https://wandb.ai/alexander-soare/lerobot/runs/5z9d8q9q/overview.

This took about 5 hours to train on an Nvida RTX H100.

Evaluation

The model was evaluated on the PushT environment from gym-pusht. There are two evaluation metrics on a per-episode basis:

• Maximum overlap with target (seen as eval/avg_max_reward in the charts above). This ranges in [0, 1].
• Success: whether or not the maximum overlap is at least 95%.

Here are the metrics for 500 episodes worth of evaluation.

The results of each of the individual rollouts may be found in eval_info.json.