Pollen Robotics Microduck: An Open-Source Biped for RL Training

Microduck is an open-source biped platform for reinforcement learning

Microduck is a 25cm tall bipedal robot designed to bridge the gap between simulation and physical deployment. It allows users to train behaviors in a physics simulator and deploy those policies directly to the hardware, making it a tool for both hobbyists and RL researchers to experiment with gait and movement policies.

Technical Specifications

Microduck is built with a focus on high-frequency control and accessibility for training. Its hardware and software stack includes:

Processor: Rockchip RK3566 with an integrated AI accelerator.
Memory & Storage: 1GB RAM and 32GB storage.
Sensors: Camera, LiDAR, and two IMUs.
Actuators: Dynamixel servos, known for higher quality and precision compared to budget alternatives.
Control Loop: 50 Hz onboard policy loop.
Connectivity: Wi-Fi, Bluetooth, and two NFC antennas.
Physicals: 800g weight and 25cm height.
Battery Life: Approximately 1 hour of runtime.

Sim-to-Real Pipeline and Software Stack

Microduck uses a simulation-first approach to behavior development. Users can train policies on their own machines or via Hugging Face Jobs, then export them to ONNX format for deployment on the physical robot.

The Training Workflow

Train in Simulation: Behaviors are learned in a physics simulation environment.
Deploy to Robot: The trained policy is transferred from simulation to the real robot.
Refine: The simulation is tuned and the policy is re-trained and re-deployed based on real-world performance.
Publish: Policies can be shared with the community.

Open Source Frameworks

The software stack is licensed under Apache-2.0 and includes the SDK, simulation environment, and the full RL training stack. Notably, the project utilizes MuJoCo (Multi-Joint dynamics with Contact), a physics engine maintained by Google DeepMind, rather than Nvidia's Isaac. Community members have noted that this choice makes the system significantly easier to set up and run on standard laptops compared to the Isaac framework.

Out-of-the-Box Capabilities

Microduck ships with seven pre-trained policies that allow it to perform specific tasks immediately:

Walking: A velocity-tracking gait.
Sitting & Standing: The ability to sit down, hold the pose, and stand back up.
Kicking: A one-shot boot followed by a return to walking.
Grabbing: Dipping the beak to the ground to scoop and pop back upright.
Roller Skating: Specialized locomotion for when roller attachments are equipped.
Self-Recovery: The ability to stand up from a flat-on-back position.

Community Insights and Analysis

Discussion among developers and robotics enthusiasts highlights several strengths and concerns regarding the Microduck platform:

Hardware and Software Trade-offs

Users have praised the use of Dynamixel servos, which are generally more expensive but offer better quality than alternatives like Feetech. However, some users expressed concern over the "open source" labeling, noting that while the software is Apache-2.0, the hardware design files (CAD/part lists) are not readily available on GitHub.

Practical Limitations

Some users pointed out potential real-world hurdles, such as the difficulty small bipedal robots face when navigating deep carpets, which can impede motion or overtax the motors.

Privacy and Safety

Because the robot includes a camera and microphones, some users raised concerns regarding privacy, particularly when using the device in children's rooms or private spaces.

Comparison to Other Platforms

Community members compared Microduck to other open-source bipeds and the Sony Aibo, noting that while Microduck provides a powerful RL training platform, it lacks the polished "cuteness" and joint geometry of high-end consumer robot pets.