Open Duck Mini

Open Duck Mini v2 is a miniature open-source biped robot inspired by the Disney BDX Droid — built under $400, standing about 42cm tall, and designed to walk using reinforcement learning policies. The project combines accessible 3D-printed hardware with a Raspberry Pi Zero 2W and trains locomotion behaviors entirely in simulation via MuJoCo Playground before deploying to the real robot.

Mechanical Design

The robot's 36 printable body parts cover the full structure: torso (trunk_top, trunk_bottom, body_front, body_back, body_middle), two 3-DOF legs (hip yaw/roll/pitch, knee, ankle pitch/roll), feet (foot_top, foot_side, foot_bottom in both PLA and TPU variants for grip), a 3-DOF articulated head (head_pitch_to_yaw, head_yaw_to_roll, head_roll_mount), LED eyes (left_eye, right_eye), antennas, speaker interface, and flash light module. The head and legs are connected via lightweight sheet structures.

Actuation

All joints use Feetech STS3215 serial bus servos, communicating at 1Mbps over a single TTL serial daisy chain. The 15-DOF configuration (3 head + 6 per leg) enables both expressive head motion and dynamic bipedal locomotion.

Embedded Runtime

The onboard computer is a Raspberry Pi Zero 2W tucked into the torso. The Open Duck Mini Runtime repository provides a Python inference stack that loads trained ONNX locomotion policies and runs them at policy rate. Pretrained walk policies are included in the main repository and can be deployed immediately after assembly.

Reinforcement Learning

The project uses sim2real transfer: policies are trained in MuJoCo Playground with domain randomization, then deployed directly to hardware. The repository includes reference motion generation tools (Open_Duck_reference_motion_generator) and actuator identification utilities based on Rhoban's BAM library.

Community

Open Duck Mini has a growing international community building and customizing their ducks, with active Discord support, a Tnkr build guide, and Chinese-language documentation. The project is maintained by @apirrone and released under Apache 2.0.