Hugging Face's Microduck: A $399 Open-Source Bipedal Robot Focused on Reinforcement Learning and Sim-to-Real Transfer

Hugging Face, a prominent organization in the open-source AI community, has introduced its latest robotics endeavor through its Bordeaux-based team, Pollen Robotics. This new offering, named Microduck, is a 25 cm tall bipedal robot that is now available for pre-order at a remarkably accessible price point of $399. Distinguishing itself from typical robotics product launches that often rely solely on demonstration videos, Microduck's core innovation lies in shipping the entire training loop. This means users are provided with the means to train the robot's movements, such as walking, sitting, kicking, roller-skating, and even recovering from falls, directly through reinforcement learning within a physics simulator. The entire process, including the training environments, reward functions, domain-randomization settings, and the crucial sim-to-real transfer recipe, is made publicly available on GitHub, fostering transparency and community-driven development. This approach represents a significant shift from Pollen Robotics' previous product, Reachy Mini, which achieved over 10,000 unit shipments. While Reachy Mini was designed for static, desk-based interactions, Microduck is engineered for dynamic, mobile tasks, emphasizing its ability to navigate and recover in real-world scenarios. Physically, Microduck stands at 25 cm tall, measures 14 cm wide, and weighs under 800 grams. Its locomotion and manipulation capabilities are powered by 15 motors strategically placed across its legs, neck, and head. A notable feature is its articulated beak, designed to pick up objects from the floor. The onboard computing is managed by a Rockchip RK3566 processor, which includes an integrated AI accelerator, complemented by 1 GB of RAM and 32 GB of storage. The sensor suite is particularly comprehensive for its price category. It includes a front-facing camera with a dedicated indicator for privacy, two Inertial Measurement Units (IMUs) – one in the body and another in the head – for motion tracking, and a compact LiDAR for range sensing, alongside an 8x8 time-of-flight matrix. Audio capabilities are provided by microphones and a speaker, and connectivity options include two NFC antennas, Wi-Fi, and Bluetooth. Power is supplied by a removable 2600 mAh NP-F550 battery, offering approximately one hour of operational time. Out of the box, Microduck comes with seven pre-trained behaviors, controllable via a bundled game controller, allowing users to experiment before delving into custom coding. These initial behaviors include walking, sitting and standing, kicking, grabbing, roller-skating, and self-recovery. Interestingly, the robot does not possess speech capabilities, but each unit generates and permanently retains its own unique audio identity upon its initial activation. The open-source nature of Microduck's training framework is a deliberate strategy by Hugging Face and Pollen Robotics to accelerate advancements in robotic locomotion and control through collaborative efforts within the AI and robotics research community.