Pollen Robotics' Microduck: a $399 open-source biped you train with RL

Pollen Robotics has opened pre-orders for Microduck, a 25 cm, 800 g bipedal robot whose every behavior is a reinforcement-learning policy the owner can retrain. It ships with seven trained moves — walking, sitting and standing, kicking, grabbing, roller skating, and self-righting from its back — but the pitch is that none of them are fixed. Owners run a loop of training a policy in physics simulation (on their own machine or via Hugging Face Jobs), deploying it to the hardware, tuning, and re-deploying, then optionally publishing the result for others to use.

The whole software stack is open under Apache-2.0 on GitHub, including the SDK, the MuJoCo-based simulation, the full RL training pipeline, and all seven shipped policies. The company leans on a sim2real approach: policies are learned against a simulated twin and moved to the real robot in one step. Hardware includes 15 motors, a camera, LiDAR, and two IMUs, with an onboard policy loop running at 50 Hz. It comes in four colorways with optional add-on packs for play gear and spare parts.

Microduck is priced at $399 (introductory, pre-tax) with delivery promised before Christmas 2026. The appeal for a technical audience is less the novelty toy and more the fully inspectable, forkable, retrainable stack — a low-cost, open platform for hands-on locomotion and RL experimentation, with a Discord community for sharing policies and troubleshooting.