GPT-6 Astra Trains Dexterous Robotic Hands via Simulation

OpenAI's GPT-6 Astra autonomously trained a dexterous robotic hand to perform tasks like pen spinning, Rubik's cube solving, and hammer use using Isaac Lab simulation. The training used 4,096 parallel simulations and took about 4.2 GPU hours. The goal is to transfer these skills to the real world.
Key insights
Autonomous Training from a Single Prompt
GPT-6 Astra was given a single prompt to train a dexterous hand to spin a pen using Isaac Lab reinforcement learning, with a Sharpa hand and self-created 3D mesh of the pen. It autonomously ran for a day and a half, including training the policy, and produced the final video. The model was allowed to search the web and download papers as needed.
Training Efficiency and Parallel Simulation
For the Rubik's cube task, GPT-6 Astra used 4,096 parallel simulations and trained for approximately 4.2 GPU hours. The zero-shot fine manipulation was already impressive, and after training it learned to spin pens, solve Rubik's cubes, and use a hammer. The training demonstrates efficient use of simulation resources.
Goal: Sim-to-Real Transfer
The stated goal is to transfer these dexterous manipulation capabilities from simulation to the real world. The training in Isaac Lab provides a scalable environment for learning complex motor skills. The results show progress toward general-purpose robotic manipulation.