Microduck open-source bipedal robot gains Isaac Lab 3.0 sim-to-real training workflow

The Microduck, an affordable open-source bipedal robot from Pollen Robotics and Hugging Face, now has a documented sim-to-real training pipeline using NVIDIA Isaac Lab 3.0. Developers can train walking policies across thousands of GPU-parallel simulations with Newton actuator modeling and domain randomization, export to ONNX, and deploy on physical hardware. The workflow includes submission to Microduck Arena on Hugging Face Hub for competitive validation.

Training pipeline and deployment details

• Isaac Lab 3.0 provides pre-configured Microduck environments with 61 observation values and 14 joint-position action outputs
• Newton with MJWarp solver and Better Actuator Models (BAM) simulates realistic servo response including firmware controller, electrical limits, friction, and delay
• Domain randomization covers mass/CoM, ground friction, joint inertia, encoder offsets, battery voltage/sag, command latency, terrain variation, and per-joint backlash
• Training reaches 1.25M steps/second with rough terrain and backlash enabled using 16 GPUs
• RSL-RL playback auto-exports both TorchScript (.pt) and ONNX (.onnx) formats; hardware deployment uses the ONNX file
• Microduck Arena accepts models with 61 observations, 14 actions, max 64 MiB size for compatibility testing and leaderboard competition

Why it matters

The integration lowers the barrier for reinforcement learning robotics research by pairing an affordable open hardware platform with a complete open-source training toolchain, potentially accelerating reproducible sim-to-real research outside well-funded labs.