Two Hours on One GPU Teaches a $399 Duck to Walk

TL;DR

Microduck is a 25cm, roughly 800g bipedal robot with 15 motors, a camera, a small LiDAR and two IMUs, and pre-orders opened on August 27 at $399 with delivery promised before Christmas. It ships with seven trained behaviors. What makes it worth your attention is not the hardware, it is that the SDK, the MuJoCo simulation environment and the full reinforcement-learning training stack are public under Apache 2.0, and the reference training run reaches a usable gait in about one to two hours on a single CUDA GPU.

What $399 actually buys

Microduck comes from Pollen Robotics, the French team Hugging Face acquired in April 2025 and the group behind the Reachy Mini desktop robot. Reachy Mini sits on your desk and moves its head. Microduck walks around your floor, which is a substantially harder control problem and the reason this release is interesting.

The spec sheet is modest and honest about it:

• 15 motors across the legs, body and an articulated beak that can grab objects.
• One camera, one small LiDAR, two IMUs. No depth array, no lidar dome, no $2,000 sensor package.
• A Rockchip RK3566 running the control stack onboard, not tethered to your laptop.
• A 50 Hz policy loop, meaning the neural network decides what all 15 joints do every 20 milliseconds.
• Seven shipped policies: walking, sitting and standing, kicking, grabbing, roller skating, and getting itself back up after a fall.

Yes, roller skating is one of the seven. Somebody at Pollen has priorities, and they are correct.

The policies are not welded shut

This is the actual news. On most consumer robots, the behaviors are a binary blob you are permitted to trigger and nothing else. Here, the seven shipped policies can be inspected, modified and retrained, and the pipeline that produced them is published in full.

The microduck repo holds the onboard side: a set of Rust daemons talking JSON-RPC over Unix sockets, where robotd owns the 50 Hz control loop and the motor bus, mediad streams the camera over WebRTC, and updaterd handles signed releases with rollback. It is a real embedded system, not a demo script. The microduck_rl repo holds the other half: the training environments, built on mjlab over MuJoCo Warp, with PPO as the learning algorithm and ONNX as the export format.

Two hours, 4,096 ducks

The training repo quotes about one to two hours for a usable gait at 4,096 parallel environments. That number is worth sitting with. You are not training one duck, you are training four thousand simulated ducks at once on a single GPU, and the policy that survives all of them is the one you flash to the real robot. If you do not have a CUDA card, the repo wires up Hugging Face Jobs so you can rent the GPU for the run instead.

The repo also ships 13 or more distinct tasks, which is roughly double the seven that come pre-loaded. Locomotion, stand-up, sit-to-stand, ball kicking, rolling and roller skating are all in there as training targets you can fork.

Sim-to-real is the whole game

A policy trained in a perfect simulator will faceplant in your living room, and the interesting engineering in microduck_rl is all about that gap. Training in an idealized sim is like learning to drive in an empty parking lot on a windless day with a brand new car: you will pass every test you set yourself and then stall at the first real hill. Domain randomization is the driving instructor who quietly lets some air out of a tire halfway through the lesson.

Concretely, the training environments randomize:

• Battery voltage and voltage sag, so the gait does not silently depend on a full charge.
• Command delay, so the policy tolerates the latency of a real control loop rather than assuming instant actuation.
• Friction, because your floor is not the sim's floor.
• Backlash of plus or minus one degree in each of the servo joints, modeling the gear play that cheap actuators actually have.

The servos are modeled with a BAM M6 actuator model for the Dynamixel XL330 units, including the voltage control law, back-EMF and friction. That is a lot more physical fidelity than most hobby robot stacks bother with, and it is the difference between a policy that works on video and a policy that works on your floor.

The price stack

The $399 is the base robot at introductory pricing, before tax and shipping. Pollen sells three optional packs on top, and the dev pack is the one to think about if you plan to actually iterate: it carries spare motors, spare batteries and NFC tags, which matters when your half-trained gait keeps driving a servo into the floor.

What to be skeptical about

Pre-orders are pre-orders. Nothing has shipped, the Christmas 2026 date is a promise rather than a track record, and $399 is described as introductory pricing, which is language that exists so it can go up later.

The software licensing is also not uniformly open. The SDK, simulation and RL stack are Apache 2.0, but the 3D models are published under Creative Commons BY-SA-NC, so the non-commercial clause applies if you were planning to print and sell your own shells.

The repos are new and small. The onboard stack sits around 428 stars and the RL stack around 76, which is a community that started this week, not a mature ecosystem with years of issues closed. If you buy one, expect to be an early adopter with all that implies.

Finally, note what a 50 Hz onboard loop on an RK3566 is and is not. It is enough to run a compact locomotion policy in real time. It is not a vision-language-action model reasoning about your kitchen, and Pollen is not claiming it is.

Why this matters beyond the duck

Robot learning has had an accessibility problem that is not really about robots. The algorithms are published, the simulators are free, and the actual barrier has been that the gap between a MuJoCo demo and a physical machine required a lab, a budget and a mechanical engineer. Microduck collapses that gap into a $399 purchase and a two-hour GPU run.

The important precedent is not the price, it is that the shipped behaviors are a starting checkpoint rather than a locked feature list. Seven policies come on the robot, and you are allowed to delete every one of them, which is more than you can say for most things you buy that have a camera in them.

Key Takeaways

• $399, pre-orders opened August 27, ships before Christmas 2026. Base robot only, before tax and shipping, at introductory pricing.
• 25cm, roughly 800g, 15 motors, one camera, one small LiDAR, two IMUs, and an articulated beak, all driven by a 50 Hz policy loop on an onboard RK3566.
• The full training stack is public Apache 2.0. mjlab over MuJoCo Warp, PPO, ONNX export, with about one to two hours to a usable gait at 4,096 parallel environments on one CUDA GPU.
• Seven shipped policies are retrainable, and the training repo carries 13 or more tasks you can fork, including the roller skating one.
• Sim-to-real is taken seriously. Domain randomization covers battery voltage, voltage sag, command delay, friction and plus or minus one degree of joint backlash, with a BAM M6 actuator model for the Dynamixel XL330 servos.
• Caveats: nothing has shipped yet, the 3D models are CC BY-SA-NC rather than Apache, and both repos are days old.

Sources: Pollen Robotics: Microduck, pollen-robotics/microduck, pollen-robotics/microduck_rl, TechCrunch, Engadget, mjlab