AI models are starting to help with more than writing robot code: they can guide simulations, watch a robot’s attempts and suggest changes to its training. In a Wired report, Will Knight used OpenAI’s GPT-6 Sol to train a virtual Microduck to walk without toppling over, while researchers describe related work with robot simulations and arms.
Watch Desk analysis
What happened
Knight says he used GPT-6 Sol with Nous Research’s Hermes agent framework and NVIDIA’s Isaac Sim. The model helped give instructions to a controller for the virtual duck, assess its progress from video and adjust its training. Knight says the work took a few hours; a similar task might once have taken weeks. The virtual duck’s vision system is not ready for a real product, and the report does not establish that its learned behaviour transfers reliably to a physical robot.
The report also describes a University of California, Berkeley lab demonstration in which OpenAI’s Astra generated a detailed robot simulation from video in about an hour, according to professor Ken Goldberg. NVIDIA researcher Jim Fan says Astra has also controlled a robot arm, including a demonstration in which it painted the Golden Gate Bridge. Separately, Fan, Goldberg and others have open-sourced Aspire, an agentic framework intended to let robots discover and use new skills.
Why it matters
Robot training often depends on detailed simulations and repeated experiments. If AI can help build those simulations, run trials and adjust training, it could make some of that work faster and more accessible to engineers. The report offers early examples, not proof that reliable physical robots can now be made with a few prompts. The duck, charming as it is, remains virtual.
Our read
The interesting shift is from AI as a coding assistant to AI as a participant in the simulation-and-training loop. That could matter well beyond pet treats, but the gap between a convincing demo and a robot that behaves safely in the real world is still the main test. Watch the tools and methods, not just the adorable hardware-shaped mascots.
What to watch
- Whether Aspire and similar frameworks make robot skills easier to reuse across tasks.
- How well simulated training transfers to physical robots outside controlled demonstrations.
- Whether researchers publish repeatable results on reliability, training time and real-world performance.
Discussion spark: Should AI-guided robot training be judged mainly by how quickly it produces a working simulation, or by whether the resulting behaviour transfers reliably to physical robots?
Sources and evidence
- Source update (11 November 2026, 19:55 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.