이 아티클은 아직 번역되지 않아 영어로 표시됩니다.
NVIDIA Microduck Explained: How an AI Agent Taught a $399 Robot Duck to Walk
How NVIDIA's $399 Microduck robot learned to walk: an AI agent on an RTX Spark laptop built the training world in 7 steps. Price, specs and the catch.
At NVIDIA's GTC Berlin keynote on October 21, 2026, Jensen Huang told four small robots to take a nap. Three dropped at once. The fourth kept walking for a few more steps before it tipped over.
The fall was funny, but it is not the story. The story is how the robots learned to walk at all: an AI agent on a laptop built a virtual world, and thousands of virtual ducks trained in it in the cloud. This guide covers what a Microduck is and the seven steps from a text prompt to a walking robot.
I have not tested one. The facts come from NVIDIA, Pollen Robotics and Hugging Face, plus Reuters, The Register and Interesting Engineering.
▶ WATCH ON YOUTUBEWatch the video
▶ WATCH ON YOUTUBEFull video
What is a Microduck?
Microduck is a small walking robot from Pollen Robotics, the robotics team at Hugging Face. Hugging Face bought the French startup in April 2025. Their first consumer robot was Reachy Mini.
- About 25 cm (10 inches) tall, under 2 lb.
- 15 motors across the legs, neck and head.
- A front camera, a depth sensor and two motion sensors.
- It walks, sits and crouches. Push it and it catches its balance. Knock it over and it gets back up.
- Its beak can pick up light objects. With wheels attached, it roller-skates.
- Seven behaviours come trained out of the box.
- The training code is open source, so owners can teach it new tricks.
Price and availability
| Item | Price |
|---|---|
| Microduck (robot, battery, USB-C cable, game controller) | $399 |
| Charger pack | $39 |
| Accessory pack | $39 |
| Developer pack | $119 |
Pre-orders opened on August 27, 2026, in Cream, Graphite, Lavender and Sky. On stage, Jensen Huang said it is already sold out. First shipments are expected by Christmas in North America and Europe.
The laptop behind it: RTX Spark
In the same keynote, Jensen Huang held up the RTX Spark. It uses a new chip, the N1X, built with Microsoft over four years, and it is now in Windows laptops and desktops. He called it "one petaflop".
RTX Spark laptops had launched the week before. The duck demo was their showcase.
His line about agents: "The way you build an agent is you ask an agent to build it for you."
From prompt to walking robot in 7 steps
NVIDIA says physical AI runs on three computers: one to train, one for the digital twin, and one inside the robot. Here is its demo, step by step.
The tools: Hermes as the agent harness, Nemotron and OpenAI's GPT-6 Astra as the models, plus FreeCAD, Blender, Isaac Sim, the Newton physics engine and Isaac Lab. Everything runs on the RTX Spark except the big jobs, which go to the cloud.
- Prompt. Describe the course and what the duck must learn. The prompt in the demo asks the agent to plan an obstacle course that teaches Microduck to walk.
- Plan. Hermes and Nemotron plan the course and list every object it needs. The agent lays the plan out as a workflow and makes a concept image before it builds anything.
- Build. GPT-6 Astra models each 3D object in FreeCAD, applies materials in Blender, and adds real physical properties with Omniverse SimReady libraries.
- Simulate. Astra puts the whole course together in Isaac Sim. It rigs the duck's joints, actuators and links, and simulates its movement with the Newton physics engine.
- Train. In the cloud, thousands of Microducks train in parallel in Isaac Lab.
- Evaluate. The policy comes back to the RTX Spark. Nemotron and Isaac Sim check it there, with many robots running at once.
- Deploy. The checked policy goes onto the real duck.
One agent ran the whole pipeline.
How the training works
Step 5 is reinforcement learning. Nobody programs each step by hand.
- Each virtual duck tries a move.
- It gets a score.
- The policy is adjusted.
- Repeat, thousands of times, across thousands of ducks.
Omniverse also simulates the duck's cameras, so it learns to see. With vision and physics together, it learns to avoid obstacles and squeeze through tight spaces. The course is made of everyday things: boxes, ramps, cables, cases and stage gear.
The point of training in simulation is simple: you crash the virtual duck, not the real one.
The honest part
- Simulation is never exactly the real world. Real floors, lights and noise are different, so robots still slip. The nap that went wrong shows this.
- The laptop did not do the heavy lifting. Training ran on NVIDIA supercomputers in the cloud. The RTX Spark planned the work and checked the result.
- It is sold out, and first units are only expected by Christmas.
Why it matters
This is NVIDIA's whole pitch for physical AI: robots that learn in a digital twin before they touch the real world. A $399 duck trained this way uses the same recipe NVIDIA is offering to every robot maker.
The same keynote announced Agile ONE SN, a reference humanoid with 31 degrees of freedom built on Isaac GR00T. Reuters reported that Germany's Agile Robots plans to produce 1,000 to 2,000 small humanoids next year with NVIDIA.
Links
- Full video: https://youtu.be/R1VtPIjb6Es
- NVIDIA GTC Berlin 2026 keynote: https://www.youtube.com/watch?v=PvFjbPYEOZI
- NVIDIA, From Prompt to Walking Robot With NVIDIA RTX Spark: https://www.youtube.com/watch?v=VvJai6HGfyE
- NVIDIA blog, keynote live updates: https://blogs.nvidia.com/blog/gtc-berlin-2026-keynote/
- Pollen Robotics, Meet Microduck: https://www.youtube.com/watch?v=reiTh7K4KSc
- The Register: https://www.theregister.com/ai-and-ml/2026/10/21/jensen-huang-sends-in-the-robot-ducks-to-sell-nvidias-ai-everywhere-vision/5304472
- Interesting Engineering: https://interestingengineering.com/ai-robotics/microduck-robots-in-nvidia-demo
- Every RTX Spark laptop compared: my guide
Would you buy one and teach it new tricks?
Not a hands-on test. Not sponsored by or affiliated with NVIDIA, Pollen Robotics or Hugging Face.


