What changed
Hugging Face and Pollen Robotics have released Microduck, an open-source bipedal robot, enabling users to train walking policies in simulation and deploy them on real hardware. The collaboration with NVIDIA's Isaac Lab and Newton allows for training, exporting policies as ONNX, and deploying them on the physical robot.
What this means in practice
Users can now train Microduck to walk using Isaac Lab's simulation framework, which supports GPU-parallel environments. The process involves installing Isaac Lab, preparing the environment, and training the robot on both flat and rough terrain. Once trained, the policy can be exported as ONNX and deployed on the real Microduck robot. The policy requires 61 observation values and outputs 14 joint-positions, with specific normalization and scaling needed for deployment.
Limitations and availability
The simulation and training require a Linux workstation with an NVIDIA GPU and driver, Git, and uv. Deployment needs a configured physical Microduck with onboard software and a supported control interface. The training process can utilize up to 16 GPUs for distributed training, achieving up to 1.25M steps per second. The policy must be exported as ONNX for deployment, and the exported files are stored in an 'exported/' directory. The Microduck Arena on Hugging Face Hub allows for policy comparison and competition, with a maximum model size of 64 MiB.
