Astra calls humanoid skills without a learned policy

A project page from Stanford and Caltech introduces HomeBody. A common humanoid stack has three layers: a vision-language model reads the scene and the instruction, a learned action model sits in the middle, and a low-level controller runs the motion. The authors ask whether that middle layer is still needed once a frontier model such as Astra is stronger. HomeBody replaces the chain. A swappable vision-language model calls a library of composable skills.

The page says that in a kitchen it has not seen before, a Unitree G1 guided by GPT Astra can tidy across the room and can fetch a remembered object from a request that does not name the place. It says this uses no training data collected for that kitchen and no extra learned policy. The Decoder carried the same system on September 27.

The robot explores first. HomeBody builds a digital twin in Nvidia Isaac Sim from the humanoid's own video, lidar, and waypoints chosen by Astra, and it keeps object locations after they leave the camera. The page shows two tasks. One gathers coffee bags in the middle and throws out specified cartons of milk and orange juice that have gone bad. In the other, the medicine starts out of view. The system finds the drawer from stored frames, retrieves the medicine, and discards the carton. The right hand opens the drawer. The left hand throws the carton out.

The vision-language model picks a skill and a target from the current view, the map, the gripper, recalled observations, and the previous result. It hands that choice over as one structured tool call, and the skill plans the motion. If a grasp closes on nothing, the pick skill can try another candidate or shift its stance. Those local retries are bounded. Past that bound, or when the failure needs a different action, the skill returns the reason and the model can move, pick a new target, or change the plan.

The page lists the limits together. Real-to-sim reconstruction adds setup time and API cost. How long a task can run is constrained by reach, manipulation, and hardware endurance, including finger servos that overheat during extended operation. GPT Astra's reasoning latency inserts pauses between skills. The local stack needs an RTX 4090 laptop GPU. The page does not give a success rate or a trial count.

The Decoder says the code is on GitHub. The repository front page says the code is coming soon. Those two statements disagree, so this piece does not treat the code as available. The project page says the robot poses in its figures are illustrative.