- Published on
A $399 Open-Source 25 cm Biped You Train with Reinforcement Learning
- Authors
- Name
- DeepHorus
Pollen Robotics, the Bordeaux-based team at Hugging Face, has announced the launch of Microduck, a $399 open-source bipedal robot that shifts the focus from pre-programmed demos to user-trainable behaviors. Priced at $399, the robot is designed for users to train its movements themselves using reinforcement learning. This marks a new approach in consumer robotics, moving beyond static functionality to interactive learning and adaptation.
Microduck follows the successful release of Reachy Mini, which has sold over 10,000 units. However, while Reachy Mini was designed as a stationary desk companion, Microduck is engineered to be mobile, capable of walking, falling, and recovering. This evolution represents a key pivot in Hugging Face's robotics strategy, emphasizing locomotion and resilience over simple interaction.
The hardware is compact and feature-rich for its price point. Standing 25 cm tall and weighing under 800 grams, it houses 15 motors, a front camera, LiDAR, two IMUs, and an AI-accelerated Rockchip RK3566 processor. A removable battery provides about one hour of operation, and it ships with seven pre-trained behaviors controllable via a bundled gamepad.
The core innovation lies in its training pipeline. All movement policies are trained in a physics simulator using PPO reinforcement learning. The process leverages a detailed BAM actuator model that accounts for real-world factors like voltage sag, friction, and gear backlash, ensuring the policies transfer effectively from simulation to the physical robot.
This approach democratizes robotics training. Pollen reports that a basic walking gait can be trained in 1–2 hours on a CUDA GPU. For users without local hardware, the same process can be run on Hugging Face Jobs. The entire software stack is Apache-2.0 licensed, though the mechanical designs remain proprietary.
Industry observers see this as a significant step in making advanced robotics accessible. By open-sourcing the training environment and reward functions, Hugging Face enables a community-driven ecosystem where users can develop and share new skills. The robot's ability to hot-swap policies mid-run highlights its flexible, modular software architecture.
Pre-orders for Microduck opened on August 27, 2026, with deliveries targeted before Christmas. The launch is backed by extensive documentation, including product pages, technical blogs, and multiple open-source repositories for the runtime and training framework. This launch could accelerate experimentation in mobile, trainable robots among developers and hobbyists alike.
