Last Update 1 hrs ago

Microduck In-Depth Analysis: Hugging Face $399 Open-Source Robot Duck

Hugging Face's Pollen Robotics has launched Microduck, a robot duck that can learn to walk on its own. At $399, it stands 9.8 inches (25 cm) tall and weighs under 1.76 lbs (800 g)—small enough to hold in one hand. It's now available for pre-order and comes with 7 pre-trained actions out of the box (walk, sit, kick, grab, roller-skate, stand up). The first batch was originally targeted for delivery before Christmas, but the official site now shows a 4-6 month lead time—ordering today basically means a spring arrival.

🔹 Hardware Specs: A Highly Integrated Research Platform

Specification	Details
Height & Weight	9.8 inches (25 cm) / about 780 g
Size / Weight	9.8 inches (25 cm) tall / <800 g
Degrees of Freedom	15 (legs, head, neck)
Brain	Rockchip RK3566 + AI accelerator, 1GB RAM / 32GB
Sensing	Front camera, 8x8 ToF LiDAR, dual IMU
Battery	NP-F550 2600mAh, about 1 hour, swappable
Actions	7 pre-trained
Training	MuJoCo + PPO, about 1-2 hours on a single GPU

Image from @pollen, copyright belongs to the original creator

🔹 Core Logic: Why a Duck Shape

Microduck is designed to traverse complex terrain and grab objects. Its roughly 9.8-inch (25 cm) duck form isn't just approachable—bipedal walking is a scaled-down challenge for physical AI balance algorithms, and its beak doubles as a jointed gripper. All 7 actions are essentially retrainable policies: train with PPO in the MuJoCo physics simulation, using a single GPU and 4,096 parallel environments, and you can train a walking gait in about 1-2 hours. No GPU? Just add a setting and run it on Hugging Face Jobs in the cloud. Export the policy to ONNX and load it into the onboard Rust runtime, and walking, standing up, and tricks can all be hot-swapped.

🔹 Evolution Path: From Virtual to Reality

Microduck supports tens of thousands of fall-training runs in the NVIDIA Isaac Sim simulator without any damage, then deploys the trained neural network policy directly to the hardware. The entire software stack is fully open source under the Apache-2.0 license.

🔹 Competitor Comparison & Buying Guide

Model	Estimated Price	Key Features
Microduck	$399	Bipedal, reinforcement learning, exceptional value
Petoi Bittle	About $289	Entry-level pricing, quadruped open-source platform
Unitree Go2	Starting around $1,600	Productivity-grade quadruped robot dog
LeRobot SO-101	About $250	Open-source robotic arm, focused on precision grasping

🔹 Pre-order Tips

The robot duck is $399 at the launch price, with four colorways (Cream/Graphite/Sky/Lavender) that share identical configurations. If you really want to train new actions, the $119 Dev pack (3 spare motors, 2 batteries, charger, HF credits) is basically a must-buy—the real entry cost is about $518 before tax.