Microduck 🤖: Tiny Robot, Big Moves! 🚀

MICRODUCK: A NEW GENERATION OF ROBOTICS
Pollen Robotics, a team originating from Hugging Face’s Bordeaux robotics lab, is pioneering a novel approach with the release of the training loop for Microduck. This innovative robot, measuring 25 cm in height, is designed to perform a wide range of complex movements—walking, sitting, kicking, roller-skating, and even recovering from falls—all through neural policies trained within a sophisticated physics simulator. These learned behaviors are then seamlessly exported to the robot’s hardware. Currently available for $399, Microduck represents a significant step forward in accessible robotics, offering a unique blend of performance and affordability. The entire training ecosystem, including environments, reward functions, and sim-to-real recipes, is openly available on GitHub, fostering collaboration and experimentation within the robotics community.

TECHNICAL SPECIFICATIONS AND SENSOR INTEGRATION
Microduck’s design incorporates a robust suite of sensors and computing power to enable its dynamic movements. The robot stands at 25 cm tall, 14 cm wide, and weighs under 800 grams. It utilizes 15 motors distributed across its legs, neck, and head, complemented by an articulated beak capable of picking up objects from the floor. Under the hood, Microduck is powered by a Rockchip RK3566 processor with an integrated AI accelerator, 1 GB of RAM, and 32 GB of storage. Notably, the sensor stack is exceptionally comprehensive for the robot's price point. A front-facing camera is strategically positioned behind a dedicated camera-use indicator. Two Inertial Measurement Units (IMUs) – one in the body and one in the head – provide precise orientation data. For range sensing, Microduck employs a compact LiDAR system featuring an 8x8 time-of-flight matrix. Furthermore, the robot incorporates microphones and a speaker, two Near Field Communication (NFC) antennas, and supports both Wi-Fi and Bluetooth connectivity. Power is supplied by a removable NP-F550 battery with a capacity of 2600 mAh, providing approximately one hour of operation.

TRAINING, CONTROL, AND KEY FEATURES
Out of the box, Microduck ships with seven pre-trained movements, controlled initially through a bundled game controller before users delve into custom coding. The robot does not possess the ability to speak, however, upon initial activation, it generates its own unique audio identity, which it retains permanently. The core training algorithms are based on ‘microduck_rl’, leveraging the mjlab (MuJoCo Warp) environment with Proximal Policy Optimization (PPO). Pollen Robotics reports that training a usable gait can take approximately one to two hours on a CUDA GPU with 4096 parallel environments. For users without local GPU access, the command “appending--hf-jobsruns” can be utilized to execute the same command on Hugging Face Jobs. The sim-to-real work relies heavily on the actuator model, incorporating a BAMM6 model of the Dynamixel XL330 for precise motor control, utilizing voltage control, back-EMF, Coulomb, Stribeck, and load-dependent friction. Per-environment randomization encompasses battery voltage, voltage sag under load, command delay, and friction magnitude, further enhancing robustness. Backlash variants train against ±1° of gear play, with a total of 2° across the 14 servo joints. The real encoder, situated on the output side of this play, reads through the observations. Trained policies are exported to ONNX with the observation normalizer integrated into the graph. Pollen Robotics strongly advises against deploying hand-converted checkpoints for this reason. Finally, a Rust runtime drives the 50 Hz control loop and the motor bus, managing the robot's operations effectively.