For years, reinforcement learning has lived behind a wall most people never get to see past — a wall built from six-figure hardware budgets, proprietary software stacks, and research lab access. Pollen Robotics thinks it has found a way to tear that wall down, and it fits in the palm of your hand.
The French robotics company best known for its open-source humanoid platform Reachy is now turning its attention to something far more attainable. Microduck, announced this week and available for pre-order on August 27, 2026 at $399, is a 25-centimeter-tall biped robot that ships with 15 motors, a built-in camera, and — critically — a complete pipeline for training AI behaviors using reinforcement learning. Units are expected to ship before Christmas 2026.
Small Robot, Big Ambitions
At 800 grams, Microduck is not going to intimidate anyone. But the hardware specs tell only part of the story. What Pollen Robotics is really selling is a workflow — one that lets hobbyists, students, and developers train robot behaviors in simulation and then push them directly onto physical hardware, a technique known as sim-to-real transfer learning.
This approach has quietly become one of the most powerful tools in professional robotics research. Rather than painstakingly programming every movement by hand, researchers train an AI agent in a virtual physics environment, letting it fail thousands of times per second at zero cost, then deploy the learned policy onto a real machine. The result is robots that can walk, recover from stumbles, and adapt to unexpected surfaces — behaviors that would take months to hard-code, learned automatically.
Pollen is making that same pipeline available to anyone willing to spend the price of a mid-range gaming peripheral.
Community Training Baked In
One of the more interesting design decisions is how tightly Microduck integrates with the broader AI community. Users can retrain behaviors either locally on their own hardware or through Hugging Face, the model-sharing platform that has become the GitHub of the machine learning world. Trained policies can then be shared back with the community — meaning a behavior someone refines in Berlin could end up running on a Microduck in São Paulo the next day.
This kind of policy-sharing ecosystem is genuinely novel at this price point. It turns what could have been a standalone toy into a node in a collaborative research network.
What You Actually Get
The base package includes the robot itself, a battery, a USB-C cable, and a game controller for direct manual operation. Pollen is also offering optional accessory packs priced between $39 and $119, likely aimed at users who want to extend Microduck’s capabilities beyond the out-of-box configuration — though specific pack contents are detailed on the Pollen Robotics product page.
Why This Moment Matters
The timing of Microduck’s announcement is not accidental. The consumer robotics space is heating up fast, with humanoid platforms dropping in price from the industrial tier toward something approaching mass-market. But most of what exists today still requires either deep pockets or deep technical expertise — often both.
Microduck positions itself differently. The target audience is not a Fortune 500 company piloting warehouse automation. It is the developer who has been running reinforcement learning experiments in simulation for two years and wants to see what happens when those policies meet the physical world. It is the university student who cannot get lab time on the department’s Boston Dynamics unit. It is the curious maker who has read every paper on locomotion control and never had a robot to test anything on.
At $399, Pollen is not just selling hardware. It is selling entry to a discipline that has historically kept most people out. Whether Microduck delivers on that promise will depend on how robust the training pipeline proves in practice — but as a statement of intent, it is hard to argue with the ambition.
