Pollen Robotics 開源 25cm、800g 強化學習雙足機械人 Microduck

Pollen Robotics Open-Sources a 25cm, 800g Reinforcement-Learning Robot Duck

Pollen Robotics, the Hugging Face-backed open-source robotics company, has released Microduck: a bipedal robot just 25cm tall and 800g in weight whose walking, recovery, and object-manipulation skills are trained largely through reinforcement learning rather than hand-coded motion. Both the robot's runtime and its RL training pipeline are now open source, making Microduck one of the smallest fully open embodied-AI development platforms on the market — and a compact reference implementation for an entire stack: MuJoCo simulation, PPO training, Sim2Real transfer, ONNX export, and real-time control in Rust.

Watch the demo (video):

https://x.com/i/status/2093175236367905091

What 25cm and 800g Can Actually Do

The numbers in the Microduck release describe a robot that punches well above its size. The hardware runs on a Rockchip RK3566 — an inexpensive ARM system-on-chip more commonly found in set-top boxes than robots — driving 15 servos at a 50Hz control frequency. That control rate matches the frequency at which the RL policies are trained, meaning the policy that emerges from simulation is deployed essentially unchanged.

The skill set, per the release details, goes well beyond basic locomotion:

The wheeled variant is a telling detail: attaching skates to the feet is detected, and the robot switches to an entirely separate RL policy for gliding, spinning, descending slopes, and rising while on skates. This is not a single model stretched to cover edge cases — it is modular policy composition, the way production robot fleets are actually built.

A Training Pipeline Built for Sim2Real

The open-sourced training project, microduck_rl, follows a pipeline that has become the industry standard for learned locomotion: MuJoCo and MuJoCo Warp for simulation, PPO as the reinforcement-learning algorithm, domain randomization for robustness, then Sim2Real transfer via ONNX export to the physical robot.

The noteworthy engineering is in the fidelity of the simulation. To close the sim-to-real gap, the training environment models phenomena that cheaper robotics projects routinely ignore:

• Motor friction
• Battery voltage variation
• Control latency
• Servo gear backlash

Gear backlash in particular — the mechanical slack between meshing gears in cheap servos — is a classic source of failed transfers, because a policy trained in an idealized simulator learns control strategies that depend on precision the hardware cannot deliver. Simulating it explicitly is the kind of unglamorous work that separates demos from deployable systems.

The Software Architecture Is the Quiet Story

The robot-side software is written mostly in Rust and deliberately structured as small, single-purpose services rather than a monolithic framework. The disclosed services include robotd (50Hz real-time control, RL policy, safety), updaterd (OTA updates with signature verification and automatic rollback), configd (Wi-Fi and device identity), btd (Bluetooth), padd (gamepad), mediad (camera, audio, WebRTC), and tofd (time-of-flight depth sensing).

Two design decisions stand out. First, services communicate over Unix Sockets with JSON-RPC 2.0, and only robotd can command motors. Phones, gamepads, and custom programs transmit only "action intents," which pass through robotd's safety checks before execution. This is a safety-oriented permission model — the agent layer can request, but only the certified control layer can act — an architecture that larger humanoid programs would do well to emulate.

Second, the OTA system implements the full enterprise update flow: download release, verify signature, switch version, restart, health check, and automatic rollback if the health check fails. A failed update bricks neither the software nor the robot.

Why It Matters

Microduck sits at the intersection of three trends reshaping embodied AI.

Cost curves. The entire compute stack — an RK3566-class SoC, 15 servos, a 50Hz control loop — is built from commodity components with no exotic hardware. When a reinforcement-learned robot can run on parts priced for consumer electronics, the floor for experimentation drops sharply, and that floor is what determines whether robotics research remains concentrated in well-funded labs or disperses to hobbyists, students, and small teams.

Open source as a distribution strategy. Hugging Face's backing of Pollen Robotics signals a bet that the path to useful robots runs through open models and open tooling, analogous to how open-weight LLMs compressed the time between research publication and production deployment. Releasing both the runtime and the training code means a researcher can reproduce the full loop — train in MuJoCo, export to ONNX, deploy on hardware — without reverse-engineering anything.

A reference platform, not just a toy. The combination of RL training, Sim2Real discipline, ONNX deployment, real-time Rust control, WebRTC media streaming, and robust OTA in a 25cm package gives Microduck outsized instructional value. It is, in effect, the architecture of a serious robot compressed into a desktop form factor. For anyone studying reinforcement learning, robot control, or embodied intelligence, the design document is the device.

The official release notes close with a disclaimer: no ducks were harmed in the making of the robot, though several were consulted. The joke lands because the underlying project is earnest — a small, cheap, thoroughly engineered platform that makes learned robot behavior something a developer can download, train, and watch stumble, recover, and roll forward on their own desk.

Sources

• Original post by @YM_AlphaNotes: https://x.com/i/status/2093175236367905091

Join GPT Brunch — AI & business community in Cyprus: https://t.me/+H1S-YWnEIxhiMmE6