Advanced robotics has traditionally been expensive and difficult to access. Research-grade robots can cost thousands—or even tens of thousands—of dollars, putting hands-on experimentation beyond the reach of many students, developers, educators, and hobbyists.
Hugging Face and its Bordeaux-based robotics team, Pollen Robotics, are taking a different approach with Microduck, a compact 25-centimeter bipedal robot designed to make physical AI and reinforcement learning more accessible.
Priced at an introductory $399 before taxes and shipping, Microduck combines a small physical platform with an open-source software ecosystem that allows developers to experiment with robotic behaviors, train reinforcement-learning policies in simulation, and deploy them on the real robot. Pre-orders opened on August 27, 2026, with first deliveries targeted before Christmas 2026.
But Microduck isn’t simply a small robot designed to sit on a desk. Its real appeal lies in the opportunity to experiment with physical AI, reinforcement learning, and sim-to-real robotics without requiring a research laboratory-sized budget.
What Is Microduck?
Microduck is a 25 cm-tall bipedal robot developed by Pollen Robotics, the Bordeaux-based robotics team that joined Hugging Face in 2025.
The robot weighs under 800 grams and has 15 motors controlling its articulated movement. It is equipped with a front camera, LiDAR, two IMUs, microphones, a speaker, and an articulated beak that can pick up objects.
Despite its compact size, Microduck can perform several physical behaviors, including walking, sitting, crouching, recovering from many common falls, and interacting with objects. It can also be equipped with roller accessories for skating.
The robot ships with seven trained behaviors, while the underlying software and training tools allow developers to retrain and create new behaviors.
Microduck Hardware at a Glance
Microduck’s hardware is designed to provide enough sensing and movement capability for experimentation with physical AI.
Key specifications include:
- Height: 25 cm
- Weight: Under 800 g
- Motors: 15
- Camera: Front-facing camera
- LiDAR: Compact time-of-flight LiDAR
- IMUs: Two
- Processor: Rockchip RK3566 with AI accelerator
- Memory: 1 GB RAM and 32 GB storage
- Connectivity: Wi-Fi and Bluetooth
- Battery: Removable rechargeable battery
- Grasping mechanism: Articulated beak
- Policy loop: 50 Hz
- Included controller: Game controller
The combination of cameras, LiDAR, motion sensors, and multiple motors gives developers a physical platform on which to experiment with perception, movement, control, and learning.
Why Reinforcement Learning Is the Big Story
The most interesting part of Microduck isn’t necessarily its appearance or size. It is the way users can train new behaviors.
Reinforcement learning (RL) is an approach in which an AI agent learns by interacting with an environment. Instead of receiving a complete set of instructions for every movement, the system takes actions and receives feedback through a reward function.
For a walking robot, for example, the training objective could reward the robot for:
- Moving forward
- Remaining upright
- Maintaining balance
- Following a desired direction
- Avoiding falls
Over many training iterations, the AI policy learns which movements produce better results.
This makes reinforcement learning particularly interesting for robotics, where manually programming every possible movement can be extremely complicated.
From Simulation to the Real Robot
Training a robot entirely on physical hardware can be slow, expensive, and potentially risky.
This is where the sim-to-real approach becomes important.
Microduck’s software ecosystem uses MuJoCo for physics simulation. Developers can train robotic policies in a virtual environment before transferring those behaviors to the physical robot. Pollen Robotics says its software stack includes the simulation and full reinforcement-learning training stack, with the robot’s policies designed to be retrainable.
The general workflow looks like this:
Design task → Create simulation → Train reinforcement-learning policy → Test in simulation → Deploy to Microduck → Evaluate real-world behavior → Retrain and improve
This creates a practical learning loop for developers who want to understand how AI moves from a virtual environment into the physical world.
Open-Source Software Makes Microduck Different
One of Microduck’s biggest advantages is its open-source software ecosystem.
Pollen Robotics states that the SDK, simulation, and full RL training stack are available as open source, allowing users to inspect, modify, fork, and retrain the software.
This is important because robotics can be difficult to learn when the underlying software is completely closed.
With an open software stack, developers can potentially:
- Study the robot’s software
- Modify existing behaviors
- Train new policies
- Experiment with reinforcement-learning algorithms
- Share improvements
- Reproduce experiments
- Build their own applications
- Learn from the wider robotics community
It is worth making an important distinction, however: Microduck’s software is open source; the robot’s mechanical and electronic design files are not described as open source by Pollen Robotics.
Who Is Microduck For?
Microduck is aimed at a broad audience interested in robotics and physical AI.
Students and Educators
For students, Microduck could provide a practical way to learn concepts that are otherwise difficult to demonstrate in a classroom.
Instead of studying reinforcement learning only through software simulations, students can potentially train policies and observe their effects on a real robot.
Educational projects could cover:
- Robotics
- Machine learning
- Reinforcement learning
- Computer vision
- Robot control
- Simulation
- AI deployment
Robotics Researchers
Researchers can use a relatively inexpensive physical platform to explore areas such as bipedal locomotion, reinforcement learning, sensor integration, and sim-to-real transfer.
The low entry price could make experimentation more accessible than purchasing conventional research robots.
AI Developers
For AI developers, Microduck provides an opportunity to move beyond purely digital models.
Developers can experiment with how AI policies behave when they have to deal with:
- Gravity
- Balance
- Physical constraints
- Sensor noise
- Hardware limitations
- Unexpected movements
- Real-world environments
This makes the platform particularly interesting as the AI industry increasingly focuses on physical AI.
Hobbyists and Makers
Microduck also has an obvious appeal to robotics enthusiasts.
The ability to purchase a ready-to-use robot and then modify its software could encourage makers to experiment with custom behaviors without building an entire robot from scratch.
What Comes With the $399 Microduck?
The introductory $399 price is for the robot package before taxes and shipping. The base package includes the robot, battery, USB-C cable, and game controller.
Pollen Robotics also offers optional accessories and development-related packages.
The company currently lists:
- Charger Pack: $39
- Dev Pack: $119
- Accessory Pack: $39
The accessory pack includes items such as rollers, a ball, laser pointer, and NFC-related accessories.
This means users can start with the base robot and add additional hardware depending on the type of experimentation they want to perform.
Microduck Is More Than a Toy
At first glance, Microduck’s duck-like design and small size make it look more like a robotic companion than a research platform.
But its underlying purpose is considerably more technical.
The robot is designed around the question of how AI can learn physical behaviors.
That makes it part of a larger movement toward physical AI, in which machine-learning models are expected to perceive and interact with the real world rather than simply process text, images, or other digital information.
Microduck provides a relatively inexpensive way to explore that transition.
Why the $399 Price Matters
The price is one of Microduck’s most important features.
Advanced robotics hardware has historically been expensive because of the cost of motors, sensors, computing hardware, mechanical components, development, and integration.
A $399 introductory robot doesn’t eliminate those engineering challenges, but it can substantially lower the barrier for people who want to experiment with physical AI.
Instead of needing access to a university robotics laboratory or a large research budget, an individual developer can potentially begin experimenting from a desk.
That could help expand the community of people working on robotics.
Building a Community Around Physical AI
Open-source software becomes even more powerful when combined with a community.
Microduck’s software can be inspected and modified, while its reinforcement-learning environment provides a foundation for experimentation.
Developers can potentially share:
- New policies
- Training methods
- Simulation environments
- Robotics projects
- Control strategies
- Tutorials
- Improvements to existing behaviors
This community-driven model could accelerate experimentation in the same way that open-source software has accelerated innovation in other areas of computing.
Microduck and the Future of Robotics
Microduck is arriving at a time when the robotics industry is moving increasingly toward AI-driven physical systems.
Large AI models have demonstrated impressive capabilities in language, vision, and multimodal reasoning. The next challenge is connecting these capabilities to machines that can physically interact with the world.
That requires solving difficult problems involving perception, planning, control, learning, and hardware.
Small platforms such as Microduck could help researchers and developers experiment with those challenges at a much lower cost.
The goal isn’t necessarily to make Microduck a replacement for larger industrial or humanoid robots. Instead, it provides an accessible environment for learning and experimenting with the technologies that may eventually power much more capable machines.
The Challenges Ahead
Microduck’s low price makes it attractive, but users should also understand its limitations.
A small research and educational robot is not the same as an industrial robot capable of performing complex manufacturing operations.
Real-world reinforcement learning remains difficult. A behavior that works perfectly in simulation may fail when transferred to physical hardware because of differences in friction, sensors, motors, mechanical tolerances, and environmental conditions.
Developers therefore need to expect experimentation, failures, and repeated training.
That is also what makes the platform valuable as a learning tool.
Summary
Hugging Face and Pollen Robotics’ Microduck represents an interesting step toward democratizing physical AI and reinforcement learning.
At an introductory price of $399, the 25 cm robot combines 15 motors, cameras, LiDAR, motion sensors, onboard computing, and an open-source software ecosystem designed for experimentation.
Its biggest contribution may not be the robot’s ability to walk, skate, pick up objects, or recover from falls. Instead, it could be the opportunity to let a much broader group of developers and students experiment with how AI learns to control physical machines.
By combining simulation, reinforcement learning, open-source software, and affordable hardware, Microduck provides a practical entry point into one of the most important areas of modern AI: teaching machines to learn and act in the physical world.
As robotics becomes increasingly connected with AI, platforms like Microduck could help turn physical AI from a technology available primarily to large research organizations into something that students, developers, researchers, and hobbyists can explore for themselves.