How an AI Agent Taught NVIDIA's $399 Robot Duck to Walk: Prompt to Policy in 7 Steps

AI-assisted disclosure: I used an AI assistant to help structure this write-up. This is **not* a hands-on test. Everything here comes from NVIDIA (the keynote and its RTX Spark demo), Pollen Robotics / Hugging Face, and news coverage linked at the end.*

At NVIDIA's GTC Berlin keynote on October 21, 2026, Jensen Huang gave four tiny robots one command: "Take a nap." Three dropped instantly. The fourth kept walking for a few steps before tipping over.

The fall is funny, but it isn't the story. The story is how they learned to walk at all: thousands of virtual copies trained in a world that an AI agent built on a laptop.

📺 Full video (8:00): https://youtu.be/R1VtPIjb6Es

What a Microduck is

• A tiny walking robot from Pollen Robotics, the robotics team Hugging Face bought in April 2025.
• About 25 cm (10 in) tall, under 2 lb.
• 15 motors (legs, neck, head), a front camera, a depth sensor and two IMUs.
• Walks, sits, crouches, catches its balance when pushed, and gets back up on its own. Seven behaviours come trained out of the box.
• $399 pre-order (opened Aug 27), now sold out. First shipments expected by Christmas in North America and Europe.
• The training code is open source, so owners can teach it new tricks.

The pipeline: prompt to walking robot in 7 steps

NVIDIA's demo ran this stack on an RTX Spark laptop (N1X chip, "one petaflop" per Jensen), bursting big jobs to the cloud:

• Agent harness: Hermes
• Models: Nemotron and OpenAI's GPT-6 Astra
• Tools: FreeCAD, Blender, Omniverse SimReady, Isaac Sim, the Newton physics engine, Isaac Lab

The steps:

1. Prompt. Describe the course and the skill: "plan an obstacle course to teach Microduck how to walk."
2. Plan. Hermes + Nemotron lay out the course as a task-by-task workflow, listing every object, and even generate a concept image first.
3. Build. GPT-6 Astra models each 3D object in FreeCAD, applies materials in Blender, and adds real physical properties from Omniverse SimReady libraries.
4. Simulate. Astra composes the course in Isaac Sim, rigs the duck's joints, actuators and links, and simulates movement with Newton.
5. Train. In the cloud, thousands of Microducks train in parallel in Isaac Lab with reinforcement learning. Each virtual duck tries a move, gets a score, and the policy is adjusted. Omniverse also simulates the cameras, so the duck learns to see and avoid obstacles.
6. Evaluate. The policy comes back to the RTX Spark, where Nemotron and Isaac Sim test it with many robots running at once.
7. Deploy. The validated policy goes onto the real duck.

Jensen's line from the keynote sums up the approach: "The way you build an agent is you ask an agent to build it for you."

Why sim-to-real matters

Pollen shows the same idea in its own footage: a move learned in simulation runs on the real robot, and when a move fails in simulation, it fails the same way on the real robot. That's the whole point. You crash the virtual duck a thousand times, not the real one.

The catch

• The nap is the honest part of the demo. A simulation is never exactly the real world. Real floors, lights and noise differ, so robots still slip up.
• The laptop didn't do the heavy lifting. Training ran on NVIDIA supercomputers in the cloud. The RTX Spark planned the work and checked the result.

Why developers should care

This is NVIDIA's "physical AI" pitch in miniature: an agent builds the world, the cloud trains the policy, a local machine evaluates it, and the robot runs it. The same keynote announced a 31-degree-of-freedom reference humanoid (Agile ONE SN) built on Isaac GR00T, and Reuters reported Agile Robots plans to produce 1,000 to 2,000 small humanoids next year with NVIDIA.

A $399 open-source duck is the cheapest way to see that recipe end to end.

Would you buy one and teach it new tricks? Tell me in the comments.

📺 Full video: https://youtu.be/R1VtPIjb6Es

Sources

• 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
• Pollen Robotics, sim to real: https://www.youtube.com/watch?v=szW7N_7B3tU
• Reuters (via KSL): https://www.ksl.com/article/51639111/nvidias-huang-hails-europes-ai-potential-flanked-by-robot-ducks
• 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