Hugging Face Unveils Microduck, a $399 Open-Source Reinforcement-Learning Robot

Hugging Face and its Pollen Robotics subsidiary unveiled Microduck on August 27, 2026, opening pre-orders for a 25 cm biped robot at an introductory price of $399 before taxes and shipping. Unlike a conventional programmable toy, Microduck executes neural movement policies onboard and ships alongside public software for simulation, reinforcement-learning training, and deployment to the physical robot.

The company is targeting first deliveries before Christmas 2026. The standard package contains the robot, a removable battery, a USB-C cable, and a game controller. Buyers can use the supplied behaviors immediately or train replacements through the project’s open-source toolchain.

Microduck can walk, sit and stand, pick up small objects with its articulated beak, kick a ball, recover from common falls, and move on optional roller attachments. The launch page says seven trained movements will be included and that each policy can be inspected, retrained, and shared.

The announcement matters less because of any single trick than because Pollen has released the operating stack and sim-to-real training system at the same time as the hardware. That gives developers a comparatively inexpensive way to change a learned behavior in simulation and then execute the resulting policy on a standardized physical machine.

1. Microduck Packages a Full Biped Platform Into 780 Grams

Microduck stands 25 cm tall and weighs approximately 780 grams. Its 15 motors provide whole-body movement, while an articulated beak serves as a small-object gripper. The compact design limits what it can carry, but it also reduces the space, reset effort, and physical risk involved in testing unstable locomotion policies.

The sensor package includes a wide-angle front camera, an 8-by-8 time-of-flight LiDAR sensor, two inertial measurement units, microphones, a speaker, and two NFC antennas located in the head and beak. Connectivity includes Wi-Fi and Bluetooth.

Onboard computing comes from a Rockchip RK3566 processor with an AI accelerator, 1 GB of RAM, and 32 GB of storage. A removable NP-F550 battery is rated for approximately one hour of operation. Pollen lists a 50 Hz onboard policy loop, meaning the runtime evaluates and applies movement commands 50 times per second.

The robot is therefore not dependent on a continuous cloud connection merely to execute a trained movement. Its public runtime loads exported ONNX policies on the local processor and drives the servos through a dedicated control loop. A gamepad is included for immediate control of the supplied behaviors.

Pollen is offering four shell colors: Cream, Graphite, Lavender, and Sky. The base package does not include the roller-skating attachments. Those are part of a separate $39 accessory pack that also includes a laser pointer, a ball, an NFC “polaroid,” and ten NFC tags. A $119 developer pack adds spare motors, motor cables, batteries, a dual charger, NFC tags, tools, screws, and Hugging Face credit.

2. The Open Stack Covers Training, Deployment, and Robot Control

Microduck’s strongest open-source claim is supported by two public GitHub repositories. The main microduck repository contains the robot runtime, SDK, control services, update system, and bundled policies. The separate microduck_rl repository contains the reinforcement-learning environments and sim-to-real pipeline.

Both repositories identify Apache 2.0 as their software license. This permissive license allows developers to inspect, modify, and redistribute the code subject to its notice and license requirements.

The robot runtime is written primarily in Rust and separates responsibilities into services. The documented architecture assigns the motor bus and policy loop to robotd, signed software installation and health-gated rollback to updaterd, Wi-Fi configuration and identity to configd, Bluetooth communication to btd, gamepad input to padd, camera streaming to mediad, and depth-sensor access to tofd. These components communicate through a common JSON-RPC contract over Unix sockets.

That architecture makes the supplied controller only one possible client. The console, mobile software, gamepad, and developer scripts can issue commands through the same interface rather than requiring separate control paths.

The project’s public material substantiates an open software stack: SDK, robot runtime, simulation, policies, reinforcement-learning code, and deployment tooling. The launch documents do not make an equally specific claim that every production hardware artifact—such as complete mechanical CAD files and electronics manufacturing files—has been released. “Open-source robot” should therefore be read most confidently as a description of the software and learning stack currently available for inspection.

3. Reinforcement Learning Runs Through a Documented Sim-to-Real Pipeline

Microduck’s movement policies are trained in a MuJoCo-based environment using proximal policy optimization, or PPO. Training runs through MuJoCo Warp, with the project requiring a CUDA-capable GPU for local training. Developers without suitable local hardware can submit the same training commands to Hugging Face Jobs.

The published quick-start documentation estimates that a usable walking gait can be trained in roughly one to two hours when running 4,096 parallel environments. That is a project estimate for the provided configuration, not a guaranteed result across every GPU or modified task.

After training, a policy is exported to ONNX. The physical robot’s runtime then executes that model at 50 Hz. The repository also provides a CPU-based MuJoCo inference script so developers can test exported policies before placing them on hardware.

The sim-to-real system attempts to model discrepancies that frequently cause simulated locomotion to fail on a physical robot. Its documented recipe includes actuator physics, domain randomization, reward-design guidance, and simulated mechanical backlash. Backlash variants represent two degrees of total gear play across each of 14 servo joints while preserving the observation and action dimensions expected by the runtime.

Published training environments cover more than ordinary walking. They include flat- and rough-terrain velocity tracking, combined walking and fall recovery, standing up from multiple initial poses, transitions between sitting and standing, ground-level beak movement, ball kicking, forward rolling, and several roller-based behaviors.

At deployment, the runtime can switch between walking, recovery, and trick policies that share a 61-dimensional observation contract. This common interface is practically important: a recovery policy can take control after a fall without requiring a completely separate robot software stack.

Pollen has also published a browser-based simulator. It runs MuJoCo compiled to WebAssembly and evaluates real ONNX policies through onnxruntime-web at 50 Hz, without a server-side simulation backend. Users can test legged and roller locomotion before obtaining the physical robot, although local browser simulation is not a substitute for validating a policy against real motors, friction, battery conditions, and sensor noise.

4. What Changes for Robotics Developers

Microduck follows Pollen’s Reachy Mini, but the two products target different development problems. Reachy Mini emphasizes conversation, perception, and expressive human-robot interaction from a desktop. Microduck is designed around locomotion and physical action: falling, recovering, carrying a small object, and transferring a simulated movement policy to real hardware.

The $399 figure lowers the hardware entry point, but it is an introductory pre-order price rather than a confirmed permanent price. Taxes and shipping are additional, and first deliveries are only targeted—not guaranteed—for before Christmas 2026.

At launch, the store lists availability in the United States, Canada, the European Union, the United Kingdom, Norway, Switzerland, Japan, and South Korea, with more countries planned later. Regional pricing can differ; the European storefront displayed a €340 price at the time of verification.

For researchers, educators, and independent developers, the practical change is the availability of a shared, purchasable body paired with reproducible policy-training code. Experiments can be exchanged as environments, rewards, checkpoints, and exported policies instead of being tied entirely to a laboratory’s custom robot.

The platform’s limitations are equally concrete. At under 800 grams and roughly 10 inches tall, Microduck is not intended for household-scale manipulation or heavy work. Its beak can retrieve small objects, but it has no arms or hands. The announced capabilities demonstrate locomotion and compact-object interaction rather than general-purpose autonomy.

Hugging Face CEO Clem Delangue described the launch as part of affordable physical AI and world-model development. The material released with Microduck, however, specifically documents reinforcement-learning policies, MuJoCo simulation, and sim-to-real deployment. It does not announce a new general-purpose world model. The immediate milestone is therefore an accessible open robotics stack, not evidence that broader physical-world intelligence has been solved.

Frequently Asked Questions

How much does Microduck cost?

The introductory pre-order price is $399 before taxes and shipping. Regional storefront prices may differ.

When will Microduck ship?

Pollen Robotics is targeting first deliveries before Christmas 2026. It is currently a pre-order product, not an immediately shipping robot.

What parts of Microduck are open source?

The SDK, robot runtime, bundled policies, simulator, reinforcement-learning environments, and sim-to-real tools are publicly available under Apache 2.0. The launch materials do not explicitly establish that every production hardware design file is available.

Is a GPU required to use the robot?

No GPU is required to run the supplied behaviors on Microduck. Training new policies locally uses a CUDA GPU, while Hugging Face Jobs is offered as a remote-training alternative.

Are the roller skates included in the $399 package?

No. The base package includes the robot, battery, USB-C cable, and game controller. Roller attachments are sold in a separate accessory pack.

Sources

• Original announcement by Clem Delangue on X
• Pollen Robotics: Meet Microduck
• Pollen Robotics: Microduck product and technical overview
• Pollen Robotics: Microduck store listing
• GitHub: Microduck robot runtime and SDK
• GitHub: Microduck reinforcement-learning environments
• Hugging Face: Microduck RL browser simulator
• Axios: Hugging Face debuts “Microduck,” a $399 robot