Stanford and Caltech researchers show a Unitree G1 doing kitchen chores from a map it built itself

Researchers at Stanford University and Caltech have published video of a Unitree G1 humanoid exploring a kitchen it had not seen before, mapping it, and then completing two multi-step chores, according to the team's HomeBody project page. In one, the robot gathered coffee bags onto an island and threw out cartons it was told had gone bad. In the other, it found medicine in a drawer it had seen during exploration, handed it to a person, and discarded a carton. The researchers say the system used no training data from that kitchen, with planning handled by a remote AI model, GPT Astra, and perception and movement run on a laptop with an RTX 4090 graphics card.

The kitchen was a test space provided by the Stanford Robotics Center, and it is the only setting shown. The team discloses that the skill videos are sped up, that drawer opening and grasping include repeated attempts, that the robot pauses while the remote model reasons, and that its finger motors overheat during extended use. The researchers describe the robot acting on its own after exploring; the page does not describe remote human control during the tasks.

What this means for buyers

This is a researcher demonstration on a robot anyone can order, not a Unitree product feature. Unitree was not involved. A buyer who orders a G1 from Unitree, which lists the base model at $13,500 before tax and shipping, gets the robot, not this software, the cloud model, or the laptop that runs it. Unitree's page lists hands and support for outside software only on the G1 EDU, priced through its sales team. The researchers do not say which configuration they used.

The demo shows work a competent housekeeper does without thinking: remembering where something was put away and going back for it. If the approach holds across more kitchens and measured trials, it likely improves what a G1 can do at kitchen tidying and fetching for an owner willing to run research software and a paid cloud model. This comes from the HomeBody team directly.

What we don't know

• How often the chores succeed. The researchers have published no success rates or trial counts, as of October 2026.
• Whether the system works in a real home. One lab kitchen is shown.
• When the code and paper will be released. The team's code repository says code is coming soon, and no paper is linked, as of October 3, 2026.

What to watch

The paper and code release. Measured success rates from more than one kitchen would be the first evidence of how often this works outside the videos.

Sources

Primary

• Gio Huh, Cayden Gu, Takara E. Truong, C. Karen Liu, Guy Tevet, HomeBody: A Humanoid That Explores, Remembers, and Acts on Its Own, The Movement Lab, Stanford University, September 2026, retrieved October 3, 2026
• Stanford-TML, homebody code repository, GitHub, retrieved October 3, 2026
• Unitree Robotics, Unitree G1 product page, undated, retrieved October 3, 2026

Spot an error? Contact corrections@cobesa.ai.