OpenAI’s GPT-6 Astra Just Hijacked a Unitree G1 to Raid Kitchens

Researchers at Stanford University and Caltech rolled out HomeBody, an architecture allowing OpenAI’s multimodal GPT-6 Astra to directly pilot a bipedal Unitree G1 humanoid inside a completely unfamiliar kitchen.

Standard robotics pipelines usually rely on an expensive middleman called Vision-Language-Action (VLA) models to translate high-level thoughts into mechanical twitches. HomeBody skips that layer entirely, letting the massive cloud model trigger a tiny library of just five built-in motor skills: navigate, pick up, put down, open drawer, and retrieve an object.

Instead of blundering into counter edges, the robot explores the space and builds a live digital twin inside Nvidia Isaac Sim using its own cameras and SLAM mapping. This gives the bot reliable spatial memory to recall where the mugs are hiding when turned away, while local visual servoing quietly fixes minor finger slips so the bot does not ping OpenAI's billing department for every millimeter adjustment.

All spatial perception and local trajectory planning run locally on a Razer Blade gaming laptop packing a mobile RTX 4090, while cognitive reasoning handles round trips to remote servers. The researchers opted not to publish explicit task completion rates or runtimes, while candidly pointing out high API latency, steep token costs, and finger servos that tend to overheat under pressure.

Tossing an unconstrained cloud-based agent into a bipedal metal skeleton inside private residences brings the industry right to the edge of an inevitable safety nightmare, especially when the creators of the underlying model keep pulling the plug on rogue autonomous behaviors.

Source: Stanford TML