GPT-6 Sol Agent Trained a Virtual Robot to Walk and Find Cats in a Few Hours

Normally, the path from the idea of "I want a robot that walks around the room and finds the cat" to a working simulation takes weeks. You have to write code, set up the environment, launch training, see why the robot falls, adjust the rewards, and repeat. Wired columnist Will Knight tried handing almost this entire loop to a language model and finished in a few hours.

What exactly the agent did

The hero of the experiment is a virtual version of Microduck, a cute robot from Pollen Robotics and Hugging Face. The author is still waiting for the physical unit, so everything happened in a simulator. The task is simple: walk around a room and recognize Bengal cats. Later the robot is supposed to dispense a treat to the author's pet, named Leona.

The toolkit: OpenAI's GPT-6 Sol, the Hermes agent harness from Nous Research, and an Nvidia DGX Spark personal computer. The model did not just write code. Through Nvidia Isaac Sim, it fed instructions to a model that controlled the virtual robot, watched videos of the results, assessed how things were going, and proposed changes to the training regime. In other words, the human acted more like a client, while the agent took over the researcher's routine work of running experiments and analyzing results.

The result: the robot learned to walk and not fall. An attempt to teach it the moonwalk from a GIF of Michael Jackson failed. After several hours of work, all that came out was a "very spirited run backwards."

What the professionals are doing

Labs are trying similar things. After the September release of the Astra model, the team of Professor Ken Goldberg at Berkeley looked at its demo of generating 3D images and asked whether it could help build simulations for training robots. Such simulations must be very accurate, otherwise complex behavior, such as a humanoid dancing or a hand grasping, transfers poorly to the real world.

One of the students gave Astra access to the Nvidia Newton tool, showed it a video clip of a real robot, and asked it to recreate the scene. According to Goldberg, in less than an hour the model produced a complete 3D simulation with tuned parameters: camera position, lighting, background, and materials.

This week, Goldberg, Nvidia researcher Jim Fan, and colleagues open-sourced the code of the Aspire framework. It allows a robot to find and apply new skills on its own. In one example, GPT-6 Astra controlled two robotic arms and solved a Rubik's cube: at first it grasped it at the wrong angle, then reasoned for a long time and still succeeded, because it accumulated skills along the way. At the CoRL conference in Austin, Fan showed another demo: the model paints the Golden Gate Bridge with a brush and paints.

Fan's idea is that an LLM now controls robots the same way it uses software tools. Models trained for physical actions (VLA) are, for it, just another tool, like an app on a computer.

Limitations and practical takeaway

Don't expect robots to start working on their own tomorrow. The author honestly admits that his duckling's vision is too weak for a real product. Reliable and efficient robot operation still requires serious engineering expertise. In addition, the demo results are described by the participants themselves, and the piece contains no independent verification.

Nevertheless, here is the takeaway for practitioners. The bottleneck in robotics is shifting from "write the code" to "formulate the task and verify the result." If you have a simulator, video analysis, and an agent capable of running experiments and analyzing outcomes, the iteration cycle shrinks from weeks to hours. For prototypes and hobby projects, this is already worth trying, and Aspire is a ready-made entry point.

Goldberg says several companies in industrial robotics and automation have already taken an interest in the approach. They see potential in manufacturing and e-commerce. Knight suggests that "vibe coding" may be followed by "vibe robotics": when programming and training robots becomes easy for both hobbyists and engineers. And Leona, if all goes well, will no longer have to go without a treat.