> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Stanford's HomeBody lets a humanoid robot explore a kitchen and tidy it on its own
- URL: https://globalfeed.ai/en/stanfords-homebody-lets-a-humanoid-robot-explore-a-kitchen-and-tidy-it-on-its-own/
- Published: 2026-09-27T22:57:31.000Z
- Updated: 2026-09-27T22:57:31.000Z
- Description: HomeBody, from The Movement Lab at Stanford, lets a vision-language model direct a humanoid robot's ready-made skills. The robot explores a kitchen it has never seen, builds a digital twin of it and cleans up from a spoken-style instruction.
- Author: GlobalFeed Editor
- Tags: Stanford, HomeBody, humanoid robots, robotics, Unitree G1, dil-en, elle, x-Stanford

Researchers from The Movement Lab at **Stanford** University and from Caltech have presented **HomeBody**, a system for humanoid robots. According to the project page, a vision-language model directs the robot's ready-made motor skills itself. The robot explores a kitchen it has never seen, builds a digital copy of the space and carries out instructions given in everyday language.

## What HomeBody is

A common setup for humanoid robots has three layers: a vision-language model that interprets images and instructions, a separately trained action model that produces commands, and a controller that executes the motion. The researchers removed the action model in the middle and tested whether the vision-language model can call reusable skills on its own. In their words, HomeBody is a system that equips frontier vision-language models with persistent spatial memory and composable humanoid skills for long tasks.

The demonstrations use a Unitree G1 humanoid and a model that the page calls GPT Astra. The researchers say the model can be swapped for another.

## How it works

Deployment takes three steps. The robot first explores the space after being told it is a kitchen robot, collecting camera images and LiDAR scans. In the second step, that data is used to build a digital twin of the room in NVIDIA Isaac Sim. The twin helps the model reason about places that the camera cannot see at that moment and return to locations it remembers. In the third step, the robot gets an everyday instruction such as "tidy up the kitchen".

The skill library has five skills: picking up an object, placing it, opening a drawer, picking an object from a drawer and walking to a location on the map. The model chooses which skill to use and on what target; planning and executing the motion is the job of the skill. If a grasp fails, the skill retries a limited number of times, and if that does not work it reports the reason to the model, which can change its plan. The skills, perception and motion planning run on a laptop with an RTX 4090 graphics card, while the vision-language model runs remotely.

## The tasks shown

The page shows two tasks. In the first, the robot gathers coffee bags on the kitchen island and throws away milk and orange juice cartons that have gone bad. In the second, it is told "I forgot my medicine, can you get it for me?"; the medicine is out of view, so the system uses stored images to find the right drawer, opens it, takes the medicine and hands it to the person. According to the researchers, no environment-specific training data was collected and no additional action policy was trained for these tasks. The system does use pretrained components for lower-body control, object segmentation and depth estimation.

The researchers also list the limits: building the digital twin adds setup time and model usage costs, the model's thinking time causes pauses between skills, and the finger motors heat up during long runs. The page gives no measured results such as success rates or trial counts; the demonstrations were filmed in a single kitchen, and the videos are sped up and include retries. No paper is linked yet, and the code repository says the code is coming soon. To see how the system does in other settings, it is worth following evaluations from independent researchers as well.

Sources: [Stanford TML: HomeBody project page](https://tml.stanford.edu/homebody/?ref=globalfeed.ai), [GitHub: Stanford-TML/homebody](https://github.com/Stanford-TML/homebody?ref=globalfeed.ai).