dp_pick_cube2
Diffusion Policy model for a robotic arm
2 min readupdated
Trained with the LeRobot framework, this model controls a robotic arm (so_follower robot type) to pick up a black cube. It uses Diffusion Policy (an approach that treats control as a generative diffusion process) to output smooth, multi-step action trajectories for contact-rich manipulation. The model takes the robot's state along with images from two cameras (overhead and wrist-mounted) and outputs 6-dimensional control actions (action). It is built for researchers and engineers working with real-world robotics on the LeRobot library.
Key features
- Dual-camera input: top-down view (
overhead) and wrist-mounted view (wrist). - Diffusion Policy architecture for stable, smooth grasping.
- Trained on a dataset of 51 episodes (22,099 frames at 30 FPS).
- Seamless integration with the LeRobot ecosystem.
What you can do
- Run the control policy on a physical
so_followerrobot to grasp objects. - Test model inference using the
lerobot-rolloutscript. - Fine-tune the model or train your own policy based on this setup.
Example
Running the policy on your robot:
lerobot-rollout \
--strategy.type=base \
--robot.type=so_follower \
--robot.port= \
--robot.cameras="{ : {type: opencv, index_or_path: , width: 640, height: 480, fps: 30}, : {type: opencv, index_or_path: , width: 640, height: 480, fps: 30}}" \
--policy.path=sudori0127/dp_pick_cube2 \
--task="Grab the black cube" \
--duration=60Read more
DetailsLicense: apache-2.0
Parameters: 0.28B
Training steps: 200,000, batch size: 32, optimizer: adam, learning rate: 0.0001
Dataset: sudori0127/pick-cube2 (51 episodes)










Comments