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

What it claims

Isaac Lab 3.0 ships ready-made Microduck walking tasks: IsaacContrib-Velocity-Flat-MicroDuck and IsaacContrib-Velocity-Rough-MicroDuck. The rough task adds stairs, slopes and uneven terrain. A policy trained on the rough task with the backlash preset was deployed on the physical robot. According to the post, it recovers its footing and resumes walking after being knocked over. With distributed training on 16 GPUs, rough terrain and backlash enabled, the authors report up to 1.25M simulation steps per second.

How it works

Simulation uses Newton with the MJWarp solver. Newton-based Better Actuator Models (BAM) model servo response. The backlash preset converts position targets into servo torques by modelling the firmware controller, motor electrical limits, friction and delay. Backlash, the play in a gearbox when a joint changes direction, is randomized independently per joint to represent uneven wear.

Per-robot domain randomization covers:

• mass and centre of mass
• ground friction
• joint inertia
• encoder offsets
• battery voltage and sag
• command latency

Training also varies terrain and applies pushes. Isaac Lab supplies the robot model, scene, observations, rewards and randomization, so no environment code is needed to start.

How to use it

Simulation and training need a Linux workstation with a supported NVIDIA GPU and driver, plus Git and uv. Deployment needs a configured Microduck with working onboard software and a control interface such as a paired gamepad. Setup uses the develop branch:

git clone --branch develop https://github.com/isaac-sim/IsaacLab.git
cd IsaacLab
uv sync
uv run isaaclab list_envs --keyword MicroDuck --show_presets

Train:

uv run isaaclab train \
  --rl_library rsl_rl \
  --task IsaacContrib-Velocity-Rough-MicroDuck \
  --num_envs 4096 \
  presets=backlash

Play back in the Newton visualizer. This step also writes exported/policy.pt (TorchScript) and exported/policy.onnx:

uv run isaaclab play \
  --rl_library rsl_rl \
  --task IsaacContrib-Velocity-Rough-MicroDuck \
  --checkpoint latest \
  --num_envs 16 \
  --viz newton_gl \
  presets=backlash

The onboard runtime loads only ONNX. Deploy with:

scp exported/policy.onnx microduck@ROBOT_HOST:/home/microduck/my_walking.onnx
ssh microduck@ROBOT_HOST
robotctl policy list
sudo robotctl policy load walk /home/microduck/my_walking.onnx
robotctl monitor
sudo robotctl policy reset walk

Microduck Arena is a Hugging Face Space with shared challenges and a leaderboard. It accepts ONNX models with 61 observation values, 14 action values and a maximum size of 64 MiB.

Limits

The post gives no training time, reward curves or quantitative sim-to-real metrics. The only hardware evidence is a video of the robot walking and recovering from falls.

The workflow depends on Isaac Lab's develop branch rather than a tagged release. Observation order, normalization and action scaling must match between training and deployment. The post defers those details to the task configuration.

The perception tasks (head-camera depth, RGB and segmentation) and the ramp traversal tasks are shown only in simulation. No hardware results are given for them.