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.
Metric | Average over 500 episodes |
---|---|
Average max. overlap ratio | 0.97 |
Success rate (%) | 71.0 |
The results of each of the individual rollouts may be found in eval_info.json.
- Downloads last month
- 105