Robotics

Microduck Robot Now Trainable in NVIDIA Isaac Lab 3.0 with Sim-to-Real Deployment

Pollen Robotics and Hugging Face enable bipedal robot training via GPU-accelerated simulation, with direct export to physical hardware.

Last verified:

Microduck Brings Sim-to-Real Bipedal Training to Developers

Pollen Robotics and Hugging Face have introduced Microduck, an affordable open-source bipedal robot that integrates with NVIDIA Isaac Lab 3.0 for end-to-end policy development. According to the Hugging Face Blog, developers can now train walking policies in GPU-accelerated simulation, export them as ONNX models, and deploy them directly on physical hardware. The framework addresses the core sim-to-real challenge: bridging the gap between frictionless virtual environments and real servo delays, mechanical play, and actuator constraints.

GPU-Accelerated Training with Domain Randomization

Isaac Lab 3.0 enables parallel policy training across thousands of simulated environments on a single GPU. According to the Hugging Face Blog, the training pipeline incorporates Newton actuator modeling and domain randomization—techniques that prepare policies for real-world variability without manual tuning. The Microduck environment provides 61 observation values (robot state, sensors) and outputs 14 joint-position commands for the bipedal gait. The framework pre-configures reward functions, scene geometry (flat and rough terrain variants), and observation normalization, reducing setup friction for new users.

Flat and Rough Terrain Tasks

The Microduck environment in Isaac Lab includes two ready-to-train variants: a flat-walking baseline and a rough-terrain task that introduces stairs, slopes, and uneven surfaces. This progression allows developers to start with simpler gaits before advancing to complex locomotion challenges. Policies trained on rough terrain are expected to generalize better to real-world deployment.

Deployment and Community Testing

Once trained, policies export as ONNX files for deterministic inference on Microduck hardware. The Hugging Face Blog indicates that developers can share trained policies on Hugging Face Hub or validate them in Microduck Arena—a competition framework for benchmarking locomotion performance. A physical Microduck requires Linux-based onboard software and a control interface (gamepad or equivalent); training requires only a supported NVIDIA GPU on a Linux workstation.

Why This Matters

This release democratizes bipedal robot training by bundling simulation, domain randomization, and export tooling into a single workflow. Teams building locomotion systems no longer need custom sim-to-real infrastructure—Isaac Lab 3.0 handles friction modeling and randomization, while ONNX standardizes deployment. For academic roboticists and hobbyists, Microduck lowers the barrier to testing walking policies on physical hardware. The Microduck Arena competition frame suggests a pathway to rapid iteration on gait design, potentially accelerating progress in bipedal control beyond research labs.

Frequently Asked Questions

What is Microduck and who builds it?

Microduck is an affordable, open-source bipedal robot developed by Pollen Robotics in collaboration with Hugging Face. It serves as a platform for training and testing walking policies in simulation before deploying them to physical hardware.

Can I train a Microduck policy without owning a robot?

Yes. Isaac Lab provides GPU-accelerated simulation environments where you can train walking policies, export them as ONNX models, and share them on Hugging Face Hub. A physical robot is only required for real-world testing.

How does the trained policy transfer from simulation to real robots?

Isaac Lab uses domain randomization, actuator modeling via Newton, and simulation of servo friction and delays to prepare policies for real-world conditions. Policies are exported as ONNX files for deterministic deployment.

What are the hardware requirements for training?

A Linux workstation with a supported NVIDIA GPU and driver, plus Git and the uv package manager. For deployment, a configured Microduck with onboard software and a control interface like a gamepad.

#sim-to-real #robotics #nvidia-isaac-lab #bipedal-locomotion #open-source