AI BriefingNVIDIA and Pollen Robotics Demonstrate Microduck Robot Locomotion Training in Isaac Lab 3.0
·2026.10.22 02:58
Key point
NVIDIA, Pollen Robotics, and Hugging Face introduce a workflow to train walking policies for the Microduck robot in Isaac Lab 3.0 and deploy them via ONNX.
Details
- NVIDIA, Pollen Robotics, and Hugging Face released a workflow to train locomotion policies for the Microduck open-source bipedal robot using Isaac Lab 3.0 and the Newton physics engine.
- Distributed training across 16 GPUs achieves up to 1.25M steps per second on rough terrain tasks with backlash enabled.
- The pipeline uses RSL-RL for reinforcement learning and exports policies as ONNX files for direct deployment on the physical robot.
- Policies are evaluated for compatibility and ranked on the Hugging Face Microduck Arena leaderboard.
Simulation and Training
- Isaac Lab provides the robot model, scene, observations, actions, rewards, and domain randomization settings.
- Newton's Better Actuator Models (BAM) simulate servo motor firmware, electrical limits, friction, latency, and mechanical backlash.
- Domain randomization includes mass, friction, joint inertia, encoder offsets, battery voltage, command latency, and terrain changes.
Deployment and Evaluation
- Trained policies are exported to ONNX format and copied to the robot via SCP.
- The robot loads the policy using robotctl commands and monitors status via robotctl monitor.
- The Microduck Arena on Hugging Face Spaces allows users to test policy compatibility and compare performance against other submissions.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.