Microduck is a small biped with big intent: make embodied AI something you can buy, train, and meaningfully improve at home. The $399 kit pairs a 25 cm, ~800 g robot (15 motors, camera, LiDAR, two IMUs) with an open SDK and a reinforcement learning stack. It ships with seven learned behaviors—walk, sit/stand, kick, grab, roller‑skate locomotion, and self‑recovery—running a 50 Hz onboard policy loop. Pre‑orders open with four colorways and optional packs (charger, accessory, developer). The core shift isn’t just price—it’s the decision to productize a sim‑to‑real pipeline that students, indie devs, and early‑stage teams can actually iterate on without a dedicated robotics lab.
The training workflow is designed to be familiar to ML practitioners: learn policies in a MuJoCo physics simulator on your machine or via hosted jobs, deploy to hardware in one step, refine the sim, and repeat. Because behaviors are first‑class policies, you can fork, retrain, and publish your improvements back to the community. That turns what used to be a fragile, lab‑specific pipeline into a standardized loop where reproducibility, environment randomization, and reward shaping are explicit levers—not hidden craft. For teams exploring locomotion, simple pick‑style moves, or curriculum learning, Microduck provides a common substrate to compare approaches without reinventing the hardware stack.
Why this matters now: embodied AI is shifting from aspiration to practice. With Microduck, a single developer can run short RL experiments, deploy them the same day, and get real‑world error signals—slip, drift, compliance, sensor noise—without lab overhead. That accelerates learning loops, improves talent pipelines, and clarifies performance bottlenecks (policy stability, terrain robustness, power draw) earlier in the project. For education, it enables hands‑on modules spanning control, perception, reward design, and sim‑to‑real gap mitigation. For founders, the calculus improves: modest capex, OSS licensing, spare parts availability, and a community policy library you can build on, not just read about.
Buyers should still plan like engineers. Expect iteration cycles that include reward re‑tuning, domain randomization, and mechanical upkeep. Surfaces, footwear (rollers vs. feet), and payloads will influence stability and battery life. The good news: the dev pack anticipates tinkering with spare motors, cables, batteries, and tools, while NFC tags and controller inputs streamline behavior switching. Treat Microduck as a testbed—log policies, track episode returns and fall rates, and be intentional about safety envelopes. With that mindset, the platform becomes a durable on‑ramp to robust sim‑to‑real practice rather than another desktop gadget.