Train in Simulation, Deploy on a Real Robot: Microduck with Isaac Lab 3.0

Pollen Robotics and Hugging Face have published a walkthrough for teaching their open-source bipedal robot, Microduck, to walk in simulation and then transferring the resulting controller to real hardware. The tutorial, released on the Hugging Face Blog, uses NVIDIA Isaac Lab 3.0 together with Newton for physics and actuator modeling. The workflow runs from installing Isaac Lab on a Linux workstation with a supported NVIDIA GPU, through training and playback, to copying an exported ONNX file onto a physical robot with scp and loading it with the robot's robotctl command-line tool.
The setup steps are intentionally short. You install uv, clone the Isaac Lab repository on its develop branch, run uv sync, and then list environments filtered by the keyword MicroDuck to confirm the robot's tasks and presets are present. Isaac Lab already ships Microduck's robot model, scene, observation and action definitions, rewards, and domain randomization settings, so no task authoring is required to get a first policy. Two variants exist: a flat-walking task and a rough-terrain task that adds stairs, slopes, and uneven ground. The authors used the rough variant for deployment.
The policy interface is fixed: 61 observation values in, 14 joint-position targets out. The tutorial stresses that observation order, normalization, and action scaling must stay consistent between training, export, and the onboard runtime, since a mismatch there is a classic cause of a policy that looks fine in the visualizer but behaves badly on the robot.
Training leans on GPU parallelism. Isaac Lab runs thousands of simulated ducks at once on a single GPU, and the example command uses 4096 environments with the RSL-RL reinforcement learning library and a backlash preset. Distributed training across multiple GPUs is also supported; the post cites a configuration with 16 GPUs that reaches up to 1.25 million steps per second with rough terrain and backlash enabled.
What makes the transfer plausible is the actuator and randomization stack. Newton's MJWarp solver and Newton-based Better Actuator Models (BAM) simulate how the servos actually respond, converting position targets into realistic torques by modeling the firmware controller, motor electrical limits, friction, and delay. Domain randomization then varies mass and center of mass, ground friction, joint inertia, encoder offsets, battery voltage and sag, and command latency per duck, along with terrain variation, applied pushes, and per-joint backlash. Backlash is the mechanical play in a gearbox: when a joint reverses direction, the transmission moves slightly before engaging. Randomizing it per joint stands in for uneven wear across a real robot's servos.
Evaluation happens in simulation first. A playback command runs the latest checkpoint, or a specified one, in the Newton visualizer with 16 environments. The RSL-RL playback backend then automatically writes two artifacts to an exported directory tied to the run: policy.pt as TorchScript and policy.onnx for the hardware handoff. A trained policy can be uploaded to the Hugging Face Hub, ideally with a model card and the training configuration so others can reproduce results.
Microduck Arena, a Hugging Face Space, is the competitive venue. It accepts policies meeting the same 61-observation, 14-action interface with a maximum file size of 64 MiB, checks compatibility, and ranks submissions on shared challenges via a leaderboard.
On the robot, deployment is a file copy plus a load command. You scp policy.onnx to the Microduck's home directory, list current policies with robotctl policy list, put the robot on its stand, and load the new file with sudo robotctl policy load walk followed by its path on the device. Reloading while the robot is already walking returns it to its home pose and resumes motion. Operators can then test standing, slow walking, turning, and stopping with the documented controls, watch state with robotctl monitor, and fall back with robotctl policy reset walk if something goes wrong. The post shows the deployed policy recovering its footing and resuming walking after being knocked over.
The final section points at extensions within Isaac Lab 3.0: harder terrain with a curriculum that raises difficulty, perception and navigation tasks that add head-camera depth and robot-state observations, dexterous manipulation via the Kuka-Allegro lift and reorient examples, and custom environments built with a project generator and populated with SimReady assets from the SimReady Central Library.
This is a tutorial rather than a product launch, and the interesting engineering choice is the emphasis on modeling what simulation usually gets wrong. Servo delay, friction, and backlash are small effects individually, but a locomotion policy exploits any discrepancy between simulator and hardware, so closing that gap with actuator models and randomized parameters is what lets a policy trained on 4096 parallel ducks survive on one real one. A likely trade-off is that heavier randomization makes training slower to converge and can produce more conservative gaits; the high step throughput reported here is partly what buys that robustness back.
The hardware requirements are worth noting for anyone planning to follow along: training needs a Linux box with a supported NVIDIA GPU, and deployment needs a configured Microduck with its onboard software running and a control interface such as a paired gamepad.
Why it matters: developers working on legged robots get an end-to-end reference path from GPU-parallel training through ONNX export to a real machine, plus a public leaderboard and model hub for comparing controllers. For teams already using Isaac Lab, Microduck is a cheap, open platform for testing sim-to-real techniques like actuator modeling and per-joint backlash randomization. It also shows the Hub and Spaces being used as shared infrastructure for robotics artifacts, not just language models, which is an inference about direction rather than a stated plan.
Based on reporting from the original publisher. Visit the source for full context and later updates.