Hugging Face has published a guide to training and deploying locomotion control for the Microduck robot
The guide from Hugging Face connects locomotion training for the Microduck robot in Isaac Lab and Newton with ONNX export for the physical robot. It includes configurations for flat and uneven terrain; simulation and training require Linux and a supported NVIDIA GPU.
Hugging Face has published a guide describing how to train locomotion control for the Microduck robot in Isaac Lab 3.0 using Newton, evaluate it in simulation and export it to ONNX for deployment on a physical robot. The control model takes 61 observation values as input and produces 14 target joint positions. The provided configurations cover flat terrain as well as stairs, slopes and uneven surfaces.
The example commands launch training with 4 096 parallel environments. The guide uses the MJWarp solver and Better Actuator Models to simulate servo motor responses, including friction, delays and mechanical backlash in the gears. According to Hugging Face, randomizing mass, center of gravity, surface friction, battery voltage and other parameters is intended to help transfer the learned control from simulation to real hardware.
Simulation and training require Linux, a supported NVIDIA GPU with the appropriate driver, and the Git and uv tools. Deployment requires a configured physical Microduck robot with working onboard software and a supported control interface. The onboard environment loads the model in ONNX format; the order of observation values, normalization and action scaling must be preserved during transfer. See the source article for details.
Why it matters
Developers of robot control systems get a concrete process from a preconfigured simulation to a file intended for a physical robot. The guide also covers modeling delays and mechanical backlash, properties of real servo motors that need to be taken into account when transferring control from simulation.
Two audiences, two different impacts
What this means
For individuals
When deploying a custom control model for the Microduck robot, all 61 input values must retain the same meaning and order, and the scaling of the 14 outputs must remain the same as during training.
For a business
Teams planning to follow this process must have a Linux workstation with a supported NVIDIA GPU for training; validating deployment also requires a physical Microduck robot with working onboard software.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.