GPT-6 Astra autonomous operation demo

Given a single prompt to perform a pen-spinning task, GPT-6 Astra autonomously ran for a day and a half, using Isaac Lab for reinforcement learning with a Sharpa hand, creating a 3D pen mesh, training the policy, and delivering a visualization video. Simulated training additionally produced skills like cube solving and hammer use.
Key insights
Single prompt enables day-long autonomous operation
GPT-6 Astra received one prompt requesting a pen-spinning task with specified techniques (Isaac Lab, Sharpa hand, 3D mesh creation). It autonomously ran for 1.5 days including training a strategy and delivering a finalized visualization video, combining web searches, paper downloads, and RL training without human intervention.
Simulation training expands dexterous capabilities
GPT-6 Astra evolves robot hand dexterity through simulated training. Zero-shot fine manipulation is already impressive, and training enables pen-spinning, cube solving, and hammer use. The cube task used 4,096 parallel simulations with approximately 4.2 GPU hours of training.
Goal: sim-to-real transfer of learned skills
The stated objective is moving these capabilities from simulated environments into the real world. The 4,096 parallel simulation approach demonstrates how scaled computation accelerates skill acquisition across multiple manipulation tasks.