Breaking

GPT-6 Astra Uses Camera Feedback to Improve a Robot's Golden Gate Bridge Painting

September 8, 2026 at 21:46 EDT

Foundation Models
AI Agents
Robotics

A recent demonstration shows GPT-6 Astra using a camera and a robot to paint a physical depiction of the Golden Gate Bridge, with its control improving across repeated attempts. The experiment is a compact example of an AI agent closing the loop between visual observation, instructions and physical action.

SEPTEMBER 2026 · OPENAI

GPT-6 Astra puts a robot to work

A physical demonstration pairs a camera, a paintbrush and a robot with repeated planning and correction to paint the Golden Gate Bridge.

THE CONTROL LOOP

From camera input to a better stroke

ObserveRead the scene through the camera.

→

PlanPlan roughly one minute of actions.

→

ActMove the robot and paint.

→

AdjustReview the result and recalibrate.

COMPARISON

OSWorld 2.0 computer-use score

OpenAI's reported score in a latency simulation; taller means higher score.

72.6%

GPT-6 Astra

65.7%

GPT-5.6 Sol

THE MODEL BEHIND THE DEMO

Built for long, tool-using tasks

1 minPlanned action horizon

97.6%FrontierMath Tier 4

$10 / $50Input / output per 1M tokens

READING THE DEMO

A compelling capability with clear boundaries

What it shows

Camera feedback and robot actions form a visible control loop.
Repeated runs improved calibration and position.

What remains unproven

The setup relied on hardware, workspace, anchors and feedback.
Reliable transfer to new tasks has not been established.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.