GPT-6 Astra added to robot arms
agents openai
| Source: HN | Original article
OpenAI's GPT‑6 Astra was tested on YAM robot arms for manipulation tasks under the Inspect Robots policy.
OpenAI has put its newest model, GPT‑6 Astra, behind the controls of a physical robot arm for the first time. In a test run reported by Robocurve on 4 September 2026, the company gave Astra command of two identical YAM manipulators operating under the Inspect Robots agent policy. The model was asked to execute a pair of simple pick‑and‑place tasks – retrieving a red block from a tabletop and depositing it into a bowl.
The experiment builds on the series of releases announced earlier in the week, when Astra was made generally available and its code‑review performance, privacy safeguards and cost profile were dissected. By moving from purely digital benchmarks to embodied manipulation, OpenAI is probing whether the model’s “world‑class” reasoning can translate into reliable actions in the real world. The company has framed the capability as a step toward the broader set of functions it associates with artificial general intelligence, noting Astra’s strong scores on cybersecurity and other benchmarks.
Why it matters is twofold. First, it demonstrates that a large language model can be integrated into a robotics control loop without bespoke firmware, potentially lowering the barrier for developers to add sophisticated perception and planning to existing hardware. Second, successful physical execution would reinforce OpenAI’s claim that Astra approaches AGI‑level versatility, a narrative that has already attracted scrutiny.
What to watch next are follow‑up trials that expand the task repertoire, involve more complex environments, or pair Astra with other robot platforms. Observers will also be keen to see whether OpenAI publishes quantitative results – success rates, latency and safety metrics – that could set a benchmark for future embodied‑AI research.
Sources
Back to AIPULSEN