Start Here: Microduck & Physical AI Resources

The 30-minute start

• Read the launch post — Meet Microduck (Pollen Robotics blog): why it exists and how it complements Reachy Mini.
• Play in your browser — open the Microduck Sandbox and the Microduck Arena on Hugging Face: the simulator and a policy-racing arena, zero installation.
• Clone the training repo (needs a CUDA GPU — or read the no-GPU note below) and train the walking policy:

git clone https://github.com/pollen-robotics/microduck_rl
cd microduck_rl
uv run train Mjlab-Velocity-Flat-MicroDuck --env.scene.num-envs 4096

No GPU? Add --hf-jobs to the same command and the training runs on Hugging Face Jobs instead.

Official code (Pollen Robotics / Hugging Face)

• pollen-robotics/microduck — the robot's brain: runtime, 50 Hz policy loop, robotctl cheat sheet, duckctl laptop control, gamepad mapping and update system. Start with its README if you own a duck.
• pollen-robotics/microduck_rl — THE training repository: mjlab (MuJoCo Warp) environments for all 13+ tasks (walking, stand-up, sit/stand, ball kick, roulade, roller skating…), PPO training, the full sim-to-real recipe (actuator model, backlash simulation, domain randomization), ONNX export, and one-command publishing of your policy to the Hugging Face Hub. Requires Python 3.12 + uv + a CUDA GPU, or HF Jobs.
• Microduck product page — official specs, accessory packs and pre-order status.
• Pollen Robotics on Hugging Face — official policies (the microduck-policies collection you can install with robotctl policy add), models and Spaces.
• Pollen Robotics on GitHub — all 227 repositories, including Microduck OS and the Reachy Mini SDK.

The training stack underneath

• mjlab — the GPU robot-learning framework microduck_rl is built on (Isaac-Lab-style API on MuJoCo Warp; needs an NVIDIA GPU for training). Its demo runs a pretrained humanoid in your browser via Colab if you just want a feel for it.
• MuJoCo Warp — the GPU physics engine underneath everything on this page.
• Hugging Face Jobs — rent the GPU instead of owning one: microduck_rl supports it natively via --hf-jobs.

Guided paths & tutorials

• Microduck RL on Jetson — Deploy the Environment (Seeed Studio Wiki) — a complete, validated setup of microduck_rl on a Jetson Orin (JetPack 7.2, CUDA 13.2, Python 3.12), drivers to first training run.
• Microduck RL on Jetson — Train and Run Official Motions (Seeed Studio Wiki) — chapter two: smoke-test training, visualise checkpoints, run the official ONNX policies, then create a custom motion of your own.
• 14 Robot Skills in 5 Days on a Single GPU (dev.to) — a practitioner's training diary on microduck_rl: the PPO settings that worked, VRAM budgeting, and why rollout GIFs beat loss curves as an acceptance test.

Learn the concepts (free)

• Hugging Face Robotics Course — the free, official course: from classical robotics to learning-based control, built around LeRobot. 30–45 minutes per unit, doable entirely in simulation.
• LearnRLWithMicroDuck — an unofficial, community-built RL course that teaches reinforcement learning on Microduck's own stack (PPO and GAE from Sutton & Barto, with labs that run real code). MIT-licensed; not affiliated with Pollen Robotics.
• Robot Learning: A Tutorial (arXiv) — the paper-length map of the whole field, from RL and behavioral cloning to generalist policies, with ready-to-run LeRobot examples.
• LeRobot — Hugging Face's robot-learning library for imitation learning (teach by demonstrating). Different technique from Microduck's reinforcement learning — knowing the difference is the fastest way to sound like you belong in this field.

Community

• Pollen Robotics Discord — where owners share policies and troubleshoot: join via the link on the official Microduck page.
• MyFirstRobotLab Experiments — our own dated lab log, once the duck lands: every training run, success and faceplant, with the exact commands we used.

Know a resource that belongs here? Tell us: myfirstrobotlab@gmail.com — this page is curated by hand and updated as the ecosystem moves.