From the source
NVIDIA's tutorial shows developers how to train a walking policy for Microduck, the open-source bipedal robot from Pollen Robotics and
Hugging Face, in Isaac Lab and Newton, then export it as ONNX and deploy it on the real robot.
- The workflow starts with installing Isaac Lab on a Linux workstation with a supported NVIDIA GPU and driver, plus Git and uv, and confirming the Microduck tasks and presets are available.
- Isaac Lab already provides Microduck's robot model, scene, observations, actions, rewards and domain randomization, with a flat-walking task and a rough variant that adds stairs, slopes and uneven terrain.
- The policy takes 61 observation values and outputs 14 joint positions, and the tutorial stresses keeping observation order, normalization and action scaling consistent between training, export and deployment.
- Training runs simulated environments in parallel on the GPU, and with 16 GPUs the configuration reaches up to 1.25M steps per second with rough terrain and backlash enabled.
- Newton's MJWarp solver and Better Actuator Models model servo response, while domain randomization varies mass, centre of mass, ground friction, joint inertia, encoder offsets, battery voltage and sag, command latency, terrain, pushes and per-joint backlash.
