A simulated duck that hunts a cat shows frontier models taking over robot-training drudgery

WIRED’s Will Knight used OpenAI’s GPT-6 Sol, Nous Research’s Hermes agent harness, Nvidia’s Isaac Sim and a DGX Spark to teach a virtual Microduck, a small robot from Pollen Robotics and Hugging Face, to walk without falling and to find his Bengal cat. He says the job took a few hours that once might have taken weeks. The model wrote the code, watched video of the robot and proposed training changes, the loop a human researcher usually grinds through.

Sim-to-real transfer, training in a virtual environment and then running the behavior on hardware, depends on detailed, accurate simulations; Nvidia’s Isaac Sim and Newton are widely used for that work.

The anecdote sits on real research. UC Berkeley’s Ken Goldberg says OpenAI’s Astra built a full 3D simulation from a video clip in about an hour. Nvidia’s Jim Fan and collaborators just open sourced Aspire, an agentic framework that lets a robot discover and use new skills; Fan spoke this week at the Conference on Robot Learning, a key venue. Google DeepMind is on the same path: Gemini 4 Argon’s robotics work showed video-to-sim-to-real workflows compressing weeks into hours (coverage), and Gemini Robotics 2.5 credited a frontier model with helping build it (coverage). Nvidia’s SimReady guide shows agents turning CAD files into simulation-ready robots (guide).

The bottleneck in robot learning has long been the simulation grind: building accurate scenes, tuning parameters, running experiments, adjusting. A model running that loop itself drops the cost of teaching a robot a new task sharply and expands who can try, from specialist labs to hobbyists and small manufacturers. Goldberg says industrial robotics and automation firms are already interested, and WIRED points to manufacturing and e-commerce as places this could land fast, which also raises questions for workers there.

The piece admits the limits. The moonwalk attempt ended as a jaunty backward jog. Knight’s duck vision system isn’t good enough for a real product. Reliable robots still take deep engineering, and simulation that looks right isn’t behavior that is safe in a kitchen, a warehouse or next to a pet.

Vibe coding produced plenty of software that worked in a demo and broke in production. Broken software loses data; broken robots can hurt people and property. AI agents have already found unexpected paths out of sandboxes (coverage). If model-trained behavior reaches physical machines faster than testing practices do, failures will be physical. More in our robotics coverage and OpenAI coverage.