Hugging Face and its French subsidiary, Pollen Robotics, have launched Microduck, a $399 open-source biped robot designed to help developers train physical AI behaviors using reinforcement learning.
Standing 25 centimeters tall, Microduck is built to navigate physical environments, waddle, crouch, sit, roller skate, recover after falling, and lift objects weighing up to 800 grams with its beak. According to Pollen Robotics, the robot comes with seven trained movements out of the box.
Developers can train new behaviors in a simulation environment and then deploy them directly to the physical hardware. The robot runs policies at 50Hz on an onboard control system powered by 15 motors. Its sensor suite includes a camera, LiDAR, and two inertial measurement units. The software development kit, simulation environment, control tools, and reinforcement-learning training stack are entirely open source.
Hugging Face CEO Clem Delangue described the device as an open-source robot that users can teach new behaviors through reinforcement learning. The company positions Microduck as a low-cost tool for developers experimenting with physical AI and models that interact with real-world spaces.
The Microduck release expands Hugging Face’s hardware footprint following its acquisition of Pollen Robotics in April 2025. The companies previously launched Reachy Mini, a desktop robot aimed at developers building AI applications on physical hardware. While Reachy Mini primarily stays on a desk and interacts with users, Pollen Robotics stated that Microduck focuses on physical mobility.
More than 10,000 units of Reachy Mini have been distributed since its launch, according to Pollen Robotics figures. The introduction of Microduck coincides with ongoing industry discussions regarding Hugging Face and a potential acquisition by Nvidia at a valuation of approximately $13 billion, though neither company has publicly confirmed a completed transaction.
Hardware and Software Specifications
Microduck integrates open-source tools with physical actuators to provide an accessible hardware platform for reinforcement learning experiments. Developers can retrain and fine-tune behaviors iteratively as they test different physical tasks.
- Dimensions: 25 centimeters tall
- Weight Capacity: Up to 800 grams lifted via beak
- Actuation: 15 onboard motors
- Perception: Camera, LiDAR, and two inertial measurement units (IMUs)
- Control Loop: 50Hz policy execution
- Price: $399 before taxes and shipping
Applications installed on the robot can still access camera or sensor feeds and transmit data depending on how they are built.