Hand me the medicine. For a human this is a trivial request; for a robot it is a chain of traps: the medicine usually isn't in view, so the robot must remember which drawer it was in, walk there, open the drawer, pick it up, and hand it over. Break any link and the task stalls halfway.
The HomeBody project from a Stanford and Caltech team recently connected a Unitree G1 humanoid robot to GPT-6 Astra and walked the robot through this entire chain in a kitchen it had never seen. The medicine was out of view, so the robot found drawer clues in previously saved keyframes, walked to the right spot, opened the drawer, retrieved the medicine, handed it to a person — and, along the way, also handled the task of discarding a carton.
🗺️ Three Steps Through an Unfamiliar Kitchen
The pipeline has three stages, each building toward that final request.
Step 1: Exploration. The robot's first instruction is simply: you are a kitchen robot, explore this space. Astra picks informative viewpoints and guides the G1's movement, logging camera frames, LiDAR scans, joint poses, and visited positions along the way. SLAM builds a map while localizing, and these observations are stored as spatial memory — the robot doesn't forget when it turns around.
Step 2: Digital twin. Astra builds a digital kitchen in Isaac Sim. The twin prioritizes accuracy over looks: SLAM-measured geometry is injected into the reconstruction process, constraining the virtual model with real measurements. The real localization map and the digital kitchen are aligned to a single coordinate system, linking camera frames, content descriptions, and spatial positions one-to-one. The drawer the robot saw earlier now has two identities: an image, and a coordinate it can return to.
Step 3: Execution. The user gives a natural-language goal, such as tidy up the kitchen, or gather the coffee bags in the middle and then discard the specified milk and orange juice cartons. Astra combines the current view and map with spatial memory, the arm's held-object state, and the previous step's result to decide which skill to call and on which target.
🔁 Failure Is Part of the Process
The system never assumes each action succeeds on the first try. A failed skill retries itself first; if retries fail, the error is reported to Astra, which either switches to another action or moves the robot and tries again.
This loop is worth a closer look. In traditional demo videos, failure means editing; in HomeBody, failure means the planning layer receives new environmental feedback and changes its mind. Compared with approaches that have a large model directly output gripper poses, here a skill library handles actions, a pretrained low-level control policy coordinates walking and manipulation, and Astra only orchestrates.
⚖️ A Different Route from End-to-End VLA
| Dimension | HomeBody-style orchestration | End-to-end VLA | |---|---|---| | Decision maker | General vision-language model | Dedicated action model | | New environment | Explore and reconstruct; no training data needed | Usually requires supplemental data and retraining | | Where actions come from | Skill library plus control policies | Model directly outputs actions | | Long-horizon tasks | Spatial memory and planning | Training distribution coverage | | Weakness | API latency at each decision step | Poor generalization to new environments |
The team states explicitly that deploying to a new kitchen requires no extra training data collection and no retraining of action policies. The weight of that claim becomes clear in the VLA comparison: over the past year, projects like DS-VLA, Holo-M, and LIFT have taken different approaches to making action models more robust, while HomeBody splits the problem between an increasingly capable general-purpose brain and a dumb-but-reliable skill library. [Inference] Every improvement in the brain makes life easier for the skill library.
🧯 Plenty of Caveats
This is a research demonstration, far from home-ready. As Robot24's report lays out plainly: the demos focus on specific tasks in a single kitchen, with generalization to other homes unverified; Astra's remote calls introduce inference latency; rebuilding the digital kitchen takes preparation time and API costs; local perception and motion planning run on an RTX 4090 laptop; and the G1's finger servos overheated during long runs. The GitHub repository says code will be released soon; external researchers cannot yet fully reproduce the work.
There's also an easily overlooked design motivation behind the spatial memory: by storing object and location information for on-demand retrieval, it avoids repeatedly feeding raw observations to the large model. Memory here is both a capability and a bill-management strategy. [Judgment]
🧭 Where's the Next Kitchen?
From tabletop arm pick-and-place to whole-room fetch tasks, the scope of LLM-controlled robots has expanded considerably in two years. What HomeBody has proven is that an orchestration architecture can skip retraining in new environments; what it has not yet proven is that cost and latency can be pushed to household-usable levels.
The test is clear: once the code is released, can a second kitchen, a second robot, and a second foundation model all get the job done without retraining? If yes, "buy a robot, then log in to a large model" starts to become a real sentence.
---
Sources: Stanford HomeBody project page and GitHub (as relayed via Robot24 and QbitAI coverage), Robot24 report (around October 2026), QbitAI piece via NetEase and Tencent News reprints, AiBoss summary.
Caveats: The project page original was not directly accessed; the three-step pipeline and failure-retry mechanism were cross-verified via Robot24 and QbitAI; limitations such as servo overheating and the RTX 4090 laptop come from Robot24; kitchen count and task list were team-chosen demos, with no success-rate statistics disclosed.