GPT-6 Astra: AI Agent Controls Robotic Hand via Reinforcement Learning

OpenAI's GPT-6 Astra was given a single prompt to make a dexterous hand spin a pen using Isaac Lab and reinforcement learning. It autonomously ran for 1.5 days, including creating a 3D mesh of the pen, searching for papers, and training the policy, resulting in a successful video demonstration.
Key insights
Autonomous Robotic Learning from a Single Prompt
GPT-6 Astra was prompted to 'make a dexterous hand spin a pen using Isaac Lab for reinforcement learning, using Sharpa hand, and create a 3D mesh of the pen. Deliver the trained policy and visualization video.' It autonomously ran for 1.5 days, including web searching, downloading papers, and training the policy. The result was a successful pen-spinning video.