XNews

GPT-6 Astra autonomous coding and dexterity demos

GPT-6 Astra autonomous coding and dexterity demos

OpenAI's GPT-6 Astra autonomously completed robotic hand twirling pen task using Isaac Lab and RL after a single prompt. It also trained to solve a Rubik's cube and use a hammer with 4,096 parallel simulations in 4.2 GPU hours. The goal is to transfer simulation capabilities to the real world.

Key insights

GPT-6 Astra autonomously designs and runs RL training for dexterity tasks

Given a prompt to make a robotic hand twirl a pen using Isaac Lab and Sharpa hand, GPT-6 Astra autonomously searched the web, downloaded papers, created 3D meshes, and ran RL training for about 1.5 days, producing a trained policy and visualization video. No human intervention was required beyond the initial prompt.

Multiple dexterity tasks achieved with efficient simulation

The model also trained policies for solving a Rubik's cube and using a hammer, demonstrating zero-shot fine manipulation after training. The Rubik's cube task used 4,096 parallel simulations consuming about 4.2 GPU hours. The trained capabilities are intended for eventual transfer from simulation to real-world robotics.

Implications for autonomous robotic skill acquisition

These results show that large language models can orchestrate complex RL training pipelines, including environment creation, data generation, and hyperparameter tuning, with minimal human guidance. This could drastically reduce the time and expertise required to develop robotic manipulation skills.

Source posts