AI agents prepare Microduck simulation
NVIDIA connects course creation, movement training and separate evaluation.
In NVIDIA’s demonstration, Hermes and Nemotron turn a prompt into an obstacle-course plan. Astra creates objects and assembles the simulated scene. Isaac Lab then improves a movement policy—a model selecting the robot’s actions—from simulated experience. Local evaluation in Isaac Sim examines a selected policy separately from learning. Developers therefore get a workflow connecting scene preparation and robot training; detailed hardware deployment lies outside this demonstration’s scope.
The linked tutorial describes that hardware handoff for walking. It requires a supported Linux GPU workstation, Git and uv; deployment also needs a configured robot with working onboard software and a control interface. Evaluate the policy in simulation first. RSL-RL playback exports an ONNX model file for the robot, alongside TorchScript. Preserve observation order, normalization and action scaling. Support the robot on its stand before loading: loading while it walks returns it to its home pose and resumes motion.
Why model the motors? The actuator documentation describes BAM, a servo model fitted to bench measurements. It represents delayed commands, supply-voltage drops and gearbox friction. This implementation requires Newton’s MJWarp solver and enabled native actuators; other backends raise an error. Its electrical supply and delay belong to a drive group, not necessarily the entire robot: separate groups can sag and respond independently.