Astra agent performs zero-shot robot control with 95% success rate in real-world tests

OpenAI’s GPT-6 Astra has crossed a threshold that robotics researchers have been chasing for years: controlling a full humanoid robot to perform pick-and-place tasks without any task-specific training. The model accomplished this using zero-shot control, meaning it figured out how to manipulate the robot’s body purely from its general training, no fine-tuning required.

What Astra actually did

In evaluations conducted by third-party organization RoboCurve, Astra achieved a 95% success rate on a physical pick-and-place task, successfully completing 19 out of 20 attempts using dual bimanual robot arms known as the I2RT YAM.

Each run consumed roughly 2.1k tokens of output and finished in approximately 2.5 minutes.

In simulation environments, the numbers got even better. Astra hit a 98% success rate, nailing 49 out of 50 single-arm assignments in RoboLab benchmarks. The model also pulled off a zero-shot cola-bottle pickup in simulated humanoid control, translating human demonstrations into robot actions using only visual inputs from cameras.

No privileged access to the robot’s internal state data. No specialized robotic datasets. Just observation-to-action loops integrating camera views and proprioception, the robot’s sense of its own body position.

Why zero-shot matters

When RoboCurve’s evaluations compared Astra against predecessor models like the Claude Fable variants, Astra outperformed them across speed, token efficiency, and reliability, particularly on gross motor tasks.

Experts involved in the evaluations flagged persistent limitations with high-precision and complex bimanual manipulation tasks. The model can grab and move objects reliably, but intricate operations that require delicate coordination between two hands still expose gaps in reliability.

From arms to full bodies

The evaluations conducted in September 2026 focused primarily on physical arm tests and embodied AI capabilities. As of that point, comprehensive documentation of Astra’s performance on full humanoid robots had yet to be revealed, with current explorations remaining focused on robotic arms and simulators.

Astra’s ability to adapt from human demonstrations to robot actions operates through an observation-to-action architecture. The model watches how a human performs a task and translates that into motor commands for a robot with a completely different physical structure, bridging that gap through its general understanding of physics and spatial reasoning rather than through hardware-specific programming.

What this means for the robotics landscape

The gap between simulation performance (98%) and real-world performance (95%) also tells an important story. That 3-percentage-point delta represents the messy reality of physical environments: unexpected lighting, imperfect surfaces, objects that don’t behave exactly like their simulated counterparts.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.