GPT-6 Astra shows promise as a robot policy, but still struggles with precision

Researchers report that GPT-6 Astra can control robots directly on a range of tasks, outperforming the other public models tested. The more useful finding is less cinematic: Astra appears capable of understanding what a robot should do, but remains unreliable when the job demands precise, dynamic or two-handed movement.

OpenAI Watch analysis

What happened

A paper on arXiv evaluates GPT-6 Astra, GPT-5.5 and DeepSeek-Flash on all 42 tasks in the RoboDojo benchmark, without task-specific fine-tuning. Across 2,100 trials, the researchers report an average success rate of 22.48% for Astra, compared with 0.88% for GPT-5.5 and 1.92% for DeepSeek-Flash.

The paper describes Astra as having a polarised capability profile. It generalised relatively well on semantic tasks, but performed poorly on precision, dynamic control and bimanual coordination. In plain English, it may understand the assignment while still making a hash of the hands-on bit. Robotics has never been especially forgiving of confident improvisation.

Why it matters

The result suggests that large language models may be useful as robot policies without first being trained specifically for every task. That could make robots easier to instruct and broaden the range of behaviours they can attempt.

But a 22.48% success rate is not a licence to send Astra into a warehouse with the keys to the forklift. The gap between recognising a goal and executing a safe, repeatable movement remains substantial. For developers, the immediate lesson is to separate high-level planning from low-level control, and to test failures on the exact hardware and environment that matter.

Our read

This is a meaningful research result, not proof that general-purpose language models have solved robotics. Astra’s lead over the comparison models is striking, while its weaknesses are just as informative. The next breakthrough may come less from making models sound cleverer and more from giving them better feedback, control loops and physical grounding.

Treat the paper as evidence that the frontier is moving, not as a deployment checklist. The robot can sometimes understand the brief. It still needs supervision when the brief involves delicate machinery, fast motion or two hands doing different things at once.

What to watch

• Whether independent teams reproduce the RoboDojo results.
• How Astra performs on physical robots rather than benchmark environments.
• Whether new control layers improve precision and dynamic movement.
• Whether OpenAI publishes broader robotics evaluations or deployment guidance.

Discussion spark: Should general-purpose AI models be allowed to control physical robots once they outperform other models on benchmarks, or should precision and failure recovery be treated as a separate bar entirely?

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.