XNews

GPT-6 Astra Autonomously Trains Robotic Hand Dexterity

GPT-6 Astra Autonomously Trains Robotic Hand Dexterity

GPT-6 Astra, given a single prompt, autonomously trained a robotic hand to perform pen spinning using Vc Lab and Sharp hand. It completed training in one and a half days, delivering a trained policy and video.

Key insights

Autonomous Training Capability

GPT-6 Astra was prompted to train a dexterous hand on pen spinning using Isaac Lab and Sharp hand. It autonomously searched the web, downloaded papers, created 3D mesh, and ran reinforcement learning for about 1.5 days, resulting in a working strategy and video. This demonstrates zero-shot planning and execution of a complex robotics task.

Training Performance and Scale

The robot hand also learned to solve Rubik's cube using 4.096 parallel simulations and about 4.2 GPU hours of training. The goal is to transfer these skills from simulation to real world.

Source posts