Hugging Face Tutorial Trains Microduck Walking Policies in Isaac Lab for Real-Robot Deployment

According to the Hugging Face blog post, the workflow is intended to bridge simulation and real hardware. A policy trained in simulation must handle the friction, delays and mechanical play of real robot servos. Isaac Lab is an open-source framework for training policies across thousands of GPU-parallel simulations; with Newton, actuator modeling and domain randomization, developers can prepare policies for real hardware.

The tutorial describes Microduck's simulation setup. Isaac Lab already provides Microduck's robot model, scene, observations, actions, rewards and domain randomization. Developers start with a flat-walking task, while a rough variant adds stairs, slopes and uneven terrain. The policy takes 61 observation values and outputs 14 joint positions. The post says observation order, normalization and action scaling must remain consistent when exporting and deploying.

For training, Isaac Lab runs simulated environments in parallel on the GPU and can distribute training across GPUs, including a configuration using 16 GPUs. In that configuration, the post says the setup can reach up to 1.25 million steps per second with rough terrain and backlash enabled. The tutorial uses Newton with the MJWarp solver and Newton-based Better Actuator Models to model servo response, alongside a backlash preset that turns position targets into servo torques by modeling the firmware controller, motor electrical limits, friction and delay.

Domain randomization is used to help the policy transfer. On each duck, the tutorial randomizes mass and center of mass, ground friction, joint inertia, encoder offsets, battery voltage and sag, and command latency. It also varies terrain, applies pushes and randomizes backlash independently per joint to represent uneven wear. The post explains that backlash is mechanical play in a gearbox, where a joint can move slightly before the transmission engages when it changes direction; varying it during training exposes the policy to differences it may encounter on the real robot.

The training examples use 4,096 simulated environments and a backlash preset, with rough terrain used for deployment; the post also offers a flat-terrain alternative. After evaluating the policy in simulation, developers export it to ONNX using Isaac Lab's RSL-RL playback workflow, which also exports a TorchScript file. The ONNX file is used with Microduck's onboard runtime.

The tutorial lists a Linux workstation with a supported NVIDIA GPU and driver, Git and uv for simulation and training. Deployment also requires a configured physical Microduck with its onboard robot software installed and working, plus a supported control interface such as a paired gamepad. Developers can share models on the Hugging Face Hub, test compatibility and compete in Microduck Arena, or run the policy on their own Microduck.

Microduck Arena is described in the post as a place to race Microduck policies on a clock you can trust. The post frames walking as the starting point and links to examples for exploring more of Isaac Lab 3.0.