Hugging Face Unveils Microduck: A $399 Open-Source 25 cm Biped You Train with Reinforcement Learning

Most robotics launches ask you to belief a demo video. Pollen Robotics, the Bordeaux robotics workforce at Hugging Face, is as a substitute delivery the coaching loop. This week it opened pre-orders for Microduck, a 25 cm bipedal robotic wherein each motion — strolling, sitting, kicking, roller-skating, standing again up after a fall — is a neural coverage skilled in a physics simulator and exported to the {hardware}. It prices $399. The coaching environments, the reward features, the domain-randomization settings, and the sim-to-real recipe are all public on GitHub. Microduck follows Reachy Mini, which has shipped greater than 10,000 models, however reverses its premise: the place Reachy Mini was constructed to sit down on a desk and work together, Microduck is constructed to go away the desk, fall over, and get again up.

The Hardware

Microduck is 25 cm tall, 14 cm vast, and below 800 g. It carries 15 motors throughout legs, neck, and head, plus an articulated beak that picks objects off the ground. Compute is a Rockchip RK3566 with an AI accelerator, 1 GB of RAM, and 32 GB of storage.

The sensor stack is unusually full for the value. A entrance digital camera sits behind a devoted camera-use indicator. Two IMUs are fitted, one within the physique and one within the head. Range sensing is a compact LiDAR, an 8×8 time-of-flight matrix. There are microphones and a speaker, two NFC antennas, plus Wi-Fi and Bluetooth. Power is a detachable NP-F550 battery, 2600 mAh, good for about an hour.

Seven skilled strikes ship within the field, pushed by a bundled sport controller earlier than you write code: stroll, sit and stand, kick, seize, roller-skate, and self-recovery. The robotic doesn’t communicate. Each unit generates its personal audio identification on first wake and retains that voice completely.

How the behaviors are literally skilled

Policies are skilled in microduck_rl, constructed on mjlab (MuJoCo Warp) with PPO. Pollen stories roughly one to 2 hours on a CUDA GPU for a usable gait at 4096 parallel environments. Without a neighborhood GPU, appending --hf-jobs runs the identical command on Hugging Face Jobs.

The sim-to-real work sits within the actuator mannequin. Each servo makes use of the BAM M6 mannequin of the Dynamixel XL330 — voltage management legislation, back-EMF, and Coulomb, Stribeck, and load-dependent friction — relatively than a perfect PD controller. Per-environment randomization covers battery voltage, voltage sag below load, command delay, and friction magnitude. Backlash variants practice towards ±1° of drugs play, 2° whole, in collection with every of the 14 servo joints within the RL format. Because the true encoder sits on the output aspect of that play, the observations learn by means of it.

Trained insurance policies export to ONNX with the statement normalizer baked into the graph. Pollen warns towards deploying hand-converted checkpoints for precisely this cause.

On the robotic, a Rust runtime drives the 50 Hz management loop and the motor bus. Every coverage shares a 61-dimensional actor statement: 48 proprioception dimensions plus instructions for twist (3), head pose (4), and physique pose (6). That shared contract is what lets stroll, recuperate, and trick insurance policies hot-swap mid-run. Environments that ignore a command slot zero-pad it relatively than dropping it.

The printed registry covers 13 duties: velocity monitoring, stand-up, sit-stand, floor decide, ball kick (70 mm, 15 g ball, actor ball-blind), roulade, and 5 roller-skating environments.

Key Takeaways

• $399 open-source-software biped, pre-orders open August 27, 2026, deliveries focused earlier than Christmas.
• 15 motors, digital camera, LiDAR, two IMUs, NFC, Wi-Fi/Bluetooth, RK3566, ~1 hour runtime.
• Policies practice in mjlab/MuJoCo Warp with PPO, ~1–2 hours for a gait at 4096 envs.
• Sim-to-real hinges on a BAM actuator mannequin plus voltage, delay, friction, and ±1° backlash randomization.
• Software is Apache-2.0; the mechanical and digital design recordsdata are usually not open.

Check out the Microduck product page, launch blog post, press kit and spec sheet, microduck runtime repo, microduck_rl training repo and announcement from Thomas Wolf. Also, be happy to comply with us on Twitter and don’t overlook to hitch our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

The publish Hugging Face Unveils Microduck: A $399 Open-Source 25 cm Biped You Train with Reinforcement Learning appeared first on MarkTechPost.