Researchers at Stanford University and the California Institute of Technology have demonstrated a Unitree G1 humanoid robot using OpenAI’s GPT-6 Astra to plan household tasks, including cleaning an unfamiliar kitchen and retrieving medicine from a drawer.
The system, called HomeBody, uses Astra as a high-level vision-language planner. Rather than generating raw motor commands, the model selects and combines reusable robot skills for navigation and manipulation, while separate control software carries out the physical movements.
The approach addresses a longstanding problem in household robotics: handling tasks that require several actions in an environment the robot has not been specifically trained for. HomeBody was designed to test whether a general-purpose vision-language model could replace a dedicated task-trained vision-language-action layer while existing robotics controllers handle execution.
Robot cleans kitchen and retrieves medicine
In one demonstration, the G1 was instructed to clean up coffee bags and throw away milk and orange juice cartons that had gone bad. The researchers describe the kitchen as an unfamiliar environment for the system, requiring the robot to identify objects, move between locations and carry out multiple actions in sequence.
A second task involved a request for medicine that was not initially visible to the robot: “I forgot my medicine, can you get it for me? Also throw out the bad carton while you are at it.” HomeBody used previously collected observations to locate the relevant drawer, navigated to it, opened the drawer with the right hand, retrieved the medicine and handed it to the person. It then discarded the carton with the left hand.
The robot first explores the environment and collects camera observations, LiDAR scans, joint information and navigation waypoints. Those observations are used in a Real2Sim process to construct a representation of the physical kitchen in NVIDIA Isaac Sim. The resulting spatial information gives Astra a reference for objects and locations that may no longer be visible from the robot’s current position.
The project uses a library of reusable skills covering navigation, picking, placing, opening a drawer and retrieving objects from a drawer. Astra determines which skill to use and supplies the relevant target, while the local system handles the detailed execution.
For example, a picking operation combines image segmentation, depth estimation, grasp prediction, motion planning, inverse kinematics and collision checking. Visual tracking systems can also help recover from unsuccessful manipulation attempts. Whole-body coordination is handled by a pretrained AMO controller rather than by Astra itself.
This division is central to HomeBody. The language-and-vision model provides task-level decisions, while established robotic systems remain responsible for turning those decisions into physical movements. The setup therefore differs from a system in which a language model directly controls the robot’s motors.
The system still has practical constraints
HomeBody depends on substantial local computing and robotics infrastructure. The physical skill execution, perception and motion planning run on a laptop equipped with an RTX 4090 GPU, while GPT-6 Astra runs remotely and communicates with the local system.
The researchers also report latency between Astra’s decisions and the resulting physical actions. Extended operation caused finger-servo overheating, and the Real2Sim preparation process adds time and API costs before a household task can begin.
The published demonstrations are not presented as a large-scale reliability benchmark. Stanford’s project page does not currently provide success-rate statistics or trial counts for the household tasks. It also notes that most skill previews are sped up to roughly seven to 10 seconds, while some drawer-opening and grasping footage contains recorded retries.
The demonstrations instead show the researchers’ intended architecture: a general-purpose model reasoning about a long sequence of physical actions while a reusable set of robot skills and lower-level controllers performs those actions.
GPT-6 Astra was introduced by OpenAI in September 2026 with vision, reasoning, computer-use and tool-calling capabilities. HomeBody applies those capabilities to a physical humanoid robot rather than a conventional software task.
The project’s GitHub repository is public, but its current documentation says the implementation code is “coming soon.”
Discover more from Aree Blog
Subscribe now to keep reading and get access to the full archive.
