Does Diffusion Policy run on ARX5?

Capabilities

Caveats

• Third-party research paper (UMI on Legs, from the Diffusion Policy authors' lab), not documented in Diffusion Policy's own docs or in LeRobot; the ARX5 LeRobot plugin documents no Diffusion Policy run.
• The ARX5 was mounted on a Unitree Go2 quadruped and driven through a learned whole-body controller; policies were trained on handheld UMI gripper demonstrations (or the public UMI cup checkpoint), not on ARX5 robot data.

Evidence 3 sources; the cell shows the highest tier

1. T3 · community reportedpapersource 2024-07-14verified 2026-10-25 (today)

   UMI on Legs, Sec. 3.3: the robot is a Unitree Go2 with a 6-DOF ARX5 arm; diffusion policy inference runs on a separate desktop RTX 4090.

   “consists of a 12-DOF Unitree Go2 quadruped robot and a 6-DOF ARX5 robot arm, both powered by the battery of Go2.”

   https://arxiv.org/abs/2407.10353v1
2. T3 · community reportedpapersource 2024-07-14verified 2026-10-25 (today)

   UMI on Legs, Fig. 3: the manipulation policy is a diffusion policy (Chi et al. default U-Net configuration) trained on real-world UMI demonstrations, whose end-effector trajectory drives the whole-body controller.

   “infers a camera-frame end-effector trajectory using a diffusion policy (a), trained using real-world UMI demonstrations.”

   https://arxiv.org/abs/2407.10353v1
3. T3 · community reportedhub-checkpointsource 2026-03-23verified 2026-10-25 (today)

   config.json type=diffusion; train_config.json dataset villekuosmanen/build_block_tower (60000 steps), whose meta/info.json robot_type=arx5 (100 episodes, 7-D joint state, front + wrist cameras). The string is arx5, not the plugin's arx5_follower; the dataset author maintains the lerobot-arx5 plugin. Auto-generated card, no results.

   https://huggingface.co/villekuosmanen/build_block_tower_diffusion