Does Diffusion Policy run on Franka Emika Panda?

Capabilities

Caveats

• The LeRobot route is community-reported only: the LeVR/LeFranX authors fine-tuned LeRobot's Diffusion Policy on a Panda (FER) with a RobotEra XHand, using the LeFranX fork, which has no Franka Hand gripper support.
• The Diffusion Policy authors' own Panda experiments (sauce pouring/spreading, bimanual tasks) ran on their own stack. The public real-stanford/diffusion_policy real-robot code targets UR5 only.

Evidence 4 sources; the cell shows the highest tier

1. T3 · community reportedpapersource 2025-09-17verified 2026-10-24 (1d ago)

   LeVR paper: ACT and Diffusion Policy as hosted by LeRobot are fine-tuned on 100 demos per task, collected with LeFranX on a Franka FER + XHand, and rolled out on the robot (10 rollouts per task, 3 tasks).

   “Using the LeRobot framework, we directly finetuned two state-of-the-art visuomotor policies, ACT and DP, on our LeFranX-collected dataset.”

   https://arxiv.org/abs/2509.14349v1
2. T1 · upstream documentedpapersource 2024-03-14verified 2026-10-24 (1d ago)

   Diffusion Policy paper (v5): real-world 6-DoF sauce pouring and periodic spreading tasks on a Franka Panda, plus bimanual tasks teleoperated onto Franka Panda arms.

   “Our Franka Panda setup and tasks are shown in Fig 10.”

   https://arxiv.org/abs/2303.04137v5
3. T1 · upstream documentedpapersource 2025-04-22verified 2026-10-24 (1d ago)

   DROID paper: all policy-learning experiments train diffusion policies (robomimic-based, released as droid_policy_learning), evaluated on 6 real tasks on the Panda-based DROID hardware.

   “To this end, we use diffusion policies [7, 16, 21, 38, 54], which leverage denoising diffusion models for action prediction”

   https://arxiv.org/abs/2403.12945v2
4. T3 · community reportedpapersource 2024-09-05verified 2026-10-24 (1d ago)

   OpenVLA paper trains Diffusion Policy from scratch on seven Franka Emika Panda tasks as a baseline. It is strong on narrow single-instruction tasks and weaker on multi-instruction ones.

   “Diffusion Policy exhibits strong performance on narrow single-instruction tasks, while Octo and OpenVLA perform better on diverse fine-tuning tasks involving multiple instructions and distractor objects.”

   https://arxiv.org/abs/2406.09246v3