Running Next-Generation LeRobot Policies using lerobot-robot-ros2-zenoh SDK

Projects
1

How can we quickly bring the latest robot learning models from the Hugging Face LeRobot ecosystem to real robots?

That question drove our latest hands-on experiment.

Using our open-source lerobot-robot-ros2-zenoh SDK, LeRobot policies connect directly to ROBOTIS hardware via a standard ROS 2 interface. This allows developers to test newly released LeRobot-native models on physical hardware right away.

In this experiment, we collected teleoperation data using CYCLO Intelligence and benchmarked four recent LeRobot policies on the ROBOTIS OMY manipulator under identical datasets and task setups:

  • GR00T N1.7

  • MolmoAct2

  • VLA-JEPA

  • FastWAM

NVIDIA Isaac GR00T N1.7 demonstrated the strongest and most stable overall performance.

Most importantly, the lerobot-robot-ros2-zenoh SDK gives us a practical framework to continuously explore rapidly evolving VLA models, world models, and policy architectures on real hardware.

This workflow extends across our entire lineup, including OMY, AI Sapiens, AI Worker, AI Manipulator, and ROBOTIS Hand. We will keep building the bridge from open robot learning models to physical testing and real-world deployment.

:clapper_board: Video: https://youtu.be/lZhRQvXYnh8
:robot: Code: https://github.com/ROBOTIS-GIT/lerobot-robot-ros2-zenoh
:memo: Docs: Running Next-Generation Policies on OMY with LeRobot | ROBOTIS Docs
:speech_balloon: Community: https://www.robotis.com/en/community.php

1 Like
ROS News for the Week of August 24th, 2026
2

Useful write-up, and the head-to-head across four LeRobot policies on one arm is the part most people cannot do, so three questions about how to read it.

How many rollouts per policy? With four policies on identical datasets and task setups, the ranking is only meaningful if the gap between adjacent policies is wider than each policy’s own run-to-run spread. At rollout counts in the tens, adjacent policies usually land inside each other’s interval.

Was the policy sampler seeded? Several of these policies sample actions, so two runs of the same checkpoint on the same episode are two draws. If the seed was not fixed, the comparison is one draw per policy rather than a measurement.

And what was the success rate with a human teleoperating the same tasks on the same cell? That is the ceiling the four policies are being compared against, and it is rarely 100 percent on contact-rich tasks.

None of this makes the result less interesting. It changes what it supports: “these four are in the same band on OMY” is a strong and useful claim, and it is a different claim from a ranking.

3

Thanks for sharing!

FYI, the git link is broken because the dashes need to be underscores. This one works: