GPT-6 Astra Controls Robot Arms, Reaching 95% on Simple Tasks

Summary

GPT-6 Astra, a general-purpose model not specifically trained for robotics, has been used to control robot arms through camera input and action planning. In one reported test, it placed a red block into a bowl in 19 of 20 attempts, a 95% success rate; in the same set of tests, it inserted a circular puzzle piece into its slot only twice, or 10%. The gap points to difficulty with precise alignment, physical contact, and force control, even when the model can bring an object close to its target. Other demonstrations include drawing the Golden Gate Bridge with an approximately $150 arm, learning a new task from a human demonstration video, and turning multi-view robot demonstrations into a simulated environment. In the described control setup, Astra interprets camera views and robot state, then outputs the end-effector position and orientation; a conventional inverse-kinematics system converts those targets into joint movements. This division lets the model plan actions without controlling every motor directly. A separate attempt to control a quadruped at 50 Hz produced about five seconds of simulated walking after 250 model calls, because inference was too slow for continuous control. The article presents this as a possible alternative to training a robotics foundation model from scratch, while noting that specialist firms pursue that route. It also describes OpenAI’s renewed interest in robotics and its stated plans to work on humanoid and other robots. The results suggest that general models may contribute perception and planning, but do not yet resolve the demands of fast feedback and precise physical interaction.