Diffusion Policy treats visuomotor control as a generative diffusion process, producing smooth, multi-step action trajectories that excel at contact-rich manipulation.

This policy has been trained and pushed to the Hub using LeRobot. See the full documentation at LeRobot Docs.

How to Get Started with the Model

For a complete walkthrough, see the training guide. Below is the short version on how to train and run inference/eval:

Train from scratch

bash

lerobot-train \
  --dataset.repo_id=${HF_USER}/<dataset> \
  --policy.type=act \
  --output_dir=outputs/train/<desired_policy_repo_id> \
  --job_name=lerobot_training \
  --policy.device=cuda \
  --policy.repo_id=${HF_USER}/<desired_policy_repo_id>
  --wandb.enable=true

Writes checkpoints to outputs/train/<desired_policy_repo_id>/checkpoints/.

Evaluate the policy/run inference

bash

lerobot-record \
  --robot.type=so100_follower \
  --dataset.repo_id=<hf_user>/eval_<dataset> \
  --policy.path=<hf_user>/<desired_policy_repo_id> \
  --episodes=10

Prefix the dataset repo with eval_ and supply --policy.path pointing to a local or hub checkpoint.

Model Details

License: apache-2.0

avyuktsachdeva/trial_move_peg_static_unity_data_converted

Model Card for diffusion

How to Get Started with the Model

Train from scratch

Evaluate the policy/run inference

Model Details

avyuktsachdeva/
trial_move_peg_static_unity_data_converted