Microduck
Microduck, a small open-source bipedal general-purpose robot released by Hugging Face in 2026: 10 inches tall, 1.7 lb, USD 400. Non-human-form (duck-like) biped.
Primary use case: open-source general-purpose bipedal robot
Height
25 cm
Weight
0.8 kg
Price
$NaN
Specifications
Physical
Height
25 cm
Weight
0.8 kg
Locomotion
bipedal
Commercial
Price
$NaN
Primary Use Case
open-source general-purpose bipedal robot
Related news
XPeng open-sourced GroundingPI , a robot perception model that beats GPT-6 Astra
The engineer who taught 1X's NEO to walk just raised $30 million to give robots a sense of touch
NVIDIA's massive robotics strategy: how all the puzzle pieces fit together
Former Tesla Optimus chief Ashish Kumar launches new industrial robot startup
Open-RAIL connects VLA models and humanoids, smoothing and speeding motion
Video Friday: Spot Plays to Show Its Whole Body Manipulation
Evan Ackerman's Video Friday roundup leads with a Robotics and AI Institute demonstration of Spot's whole-body manipulation, noting that direct torso involvement is hard to discern. It also collects videos of ETH Zurich humanoid locomotion research, Reachy Mini racing Microduck, monkeyBot 2.0 pipe climbing, ANYmal X, a LimX Dynamics tennis robot, DEEPRobotics machines, MonoRace drone racing, and the amphibious SAWbot; no funding is discussed.
Jensen Huang’s Robot Duck Somersault Shows RTX Spark’s Physical AI Push
At the GTC Berlin keynote on 21 October 2026, NVIDIA CEO Jensen Huang asked Pollen Robotics' small open-source Microduck bipeds to somersault, with mixed results. The demonstration highlighted NVIDIA's three-computer physical AI workflow: RTX Spark laptops running AI agents that set up Omniverse digital twins, cloud training of thousands of simulated Microducks with Isaac Lab, and on-robot inference, showing that accessible hardware can orchestrate the full loop from simulation to physical deployment.
Using AI Agents to Teach Robots Like Microduck to Walk in Simulation | NVIDIA Technical Blog
NVIDIA describes a workflow in which Hermes, Nemotron 3.5 Bolt, and Astra plan an obstacle course, create simulation assets, and assemble the environment for the Microduck robot. The robot's perception and locomotion policy is trained in Isaac Lab with 4,096 parallel environments and an NVIDIA RTX PRO 6000 Blackwell Workstation Edition GPU, then evaluated locally in Isaac Sim on RTX Spark. The workflow also covers preparing for physical deployment through better actuator modeling, domain randomization, and hardware testing.