A useful test of embodied AI is whether a learned behavior survives contact with real motors, sensors, and falls—not just whether it looks convincing in simulation. Microduck makes that gap the central engineering problem: policies are trained in simulation, then run on a compact physical robot through an onboard control system.
What Sets It Apart
- A complete path from learned policy to physical action. The runtime loads ONNX policies trained with MuJoCo and PPO, connecting simulation work to a real robot rather than stopping at a training demo.
- A real-time control loop on modest hardware. A Rockchip RK3566 runs the control loop at 50 Hz and drives 15 servos, making the project a practical example of deploying learned behaviors at the edge.
- Multiple behaviors share a policy interface. Walking, recovery, and other behaviors can be switched on the robot, while the companion training project provides the simulation and sim-to-real workflow.
- The robot is also a software platform. Rust daemons manage motors, connectivity, camera streaming, and reversible software updates, so the learning system is part of a maintainable robot rather than a standalone model.
Who It’s For
Great fit for robotics developers and researchers exploring reinforcement learning, sim-to-real transfer, and small-scale embodied systems. Look elsewhere if you need a general-purpose humanoid, a turnkey autonomous robot, or a training stack that works without compatible hardware and a CUDA-capable GPU.
