HomeBody: Stanford Lets GPT-6 Astra Drive a Humanoid Without a VLA in Between

Most humanoid companies now build robot “brains” in three layers. A large vision-language model (VLM) understands the scene and the instruction. A trained vision-language-action model (VLA) turns that understanding into motor commands. A fast low-level controller keeps the robot balanced and its joints on target. Figure’s Helix popularised this split in early 2025 with a slow “System 2” VLM running at 7–9 Hz and a fast “System 1” policy at 200 Hz. Helix 02 later added a 1 kHz “System 0” for whole-body balance.

A project page published in late September by Stanford’s Movement Lab and Caltech asks a blunt question: now that frontier models are this capable, is the middle layer still needed? Their system, HomeBody, removes it. GPT-6 Astra, OpenAI’s flagship model released on 3 September, calls a small library of motor skills directly through structured tool calls. Using that setup, a Unitree G1 humanoid cleaned up a kitchen it had never been in and fetched a remembered object from a drawer. According to the authors, it needed no environment-specific training data and no additional policy learning.

Who and what

The authors are Gio Huh (Caltech, a visiting undergraduate researcher on a SURF fellowship), Cayden Gu and Takara E. Truong (Stanford). C. Karen Liu and Guy Tevet of Stanford advised jointly. The work was carried out in a kitchen at the Stanford Robotics Center. The team published a project page dated September 2026, and a GitHub repository was created on 26 September. As of today the README says “Code coming soon”, so some coverage describing the code as already available is premature. There is also no arXiv paper yet. The BibTeX entry points only to the project page.

The page shows two tasks:

“Tidy the kitchen”: gather all the coffee bags on the island and throw away the spoiled milk and orange-juice cartons. This takes repeated trips, grasps and placements across the room.
“I forgot my medicine, can you get it for me?”: the medicine starts out of view inside a drawer. The robot uses stored keyframes to work out which drawer, opens it with its right hand, retrieves the pill bottle, hands it over, then throws out a carton with its left hand.

How it works

The pipeline has three stages.

1. Explore. The robot is told “You are a kitchen robot, please explore the space!” Astra chooses the waypoints. The robot records its own RealSense D435i camera stream, LiDAR scans with SLAM, joint poses and waypoints. A human-shot 0.5× iPhone video adds visual detail.

2. Real2Sim. Astra then acts as a “Real2Sim agent”: it builds a digital twin of the kitchen in NVIDIA Isaac Sim from that data. The authors stress that video alone is not enough. Working from appearance only, the agent has to guess room dimensions, so the measured SLAM geometry is what makes the twin accurate enough to navigate and reach into. The robot localises itself with Super Odometry, and ICP registration aligns its map with the simulation. Ego-camera observations are stored in the same coordinate frame with text descriptions. That persistent spatial memory lets it find things it can no longer see.

3. Act. At each step Astra receives the current ego view, the map context, the gripper state, recalled observations and the result of the previous skill. It picks one of five skills (navigate, pick, place, open drawer, pick from drawer) and a target. The interfaces are deliberately simple:

Pick takes an image point normalised to 0–1000 and which hand to use. That point prompts SAM 2 segmentation, Fast-FoundationStereo estimates depth, and the grasp is computed analytically. The arm follows a minimum-jerk spline, with inverse kinematics solved along the path and a collision check.
Navigate takes a 2D goal and a facing point, in metres, in map coordinates.
Place takes a hand, a 3D release point in the torso frame and a release distance.
Open drawer bundles handle alignment, a hooking posture and walking backwards into one skill.

Balance while reaching comes from the pretrained AMO whole-body controller. Arm and hand commands run at 250 Hz, and the AMO policy updates at 50 Hz. Everything except the model runs on one Razer Blade laptop with an RTX 4090. Astra runs remotely in the cloud.

Error handling is where it looks most like a software agent. SAMURAI tracking and visual servoing correct small misalignments without asking the model. If a grasp closes on nothing, the skill retries a bounded number of times with a different candidate or stance. Only when local recovery runs out does the skill return a failure reason to Astra, which can reposition, pick a new target or replan. For developers, this is the familiar tool-calling agent loop, with motors as the side effects.

Why it matters

The industry consensus of the past two years is that general-purpose robots need VLAs trained on huge amounts of robot data. That belief is why labs are building teleoperation farms and why Figure’s “Project Go-Big” aims to learn from human video at internet scale. HomeBody is a data point in the other direction: with good spatial memory and a well-designed skill API, a general-purpose model with no robot-specific training can handle long-horizon planning. That changes where the engineering effort goes. Instead of collecting demonstrations, you build reliable skills with clean interfaces, which is closer to building an SDK than to training a policy. The skill library is explicitly pluggable: “connect a learned policy, a classical algorithm or another controller through the shared interface.”

It also fits a larger shift. In September Sam Altman said on the Sources podcast that OpenAI “will definitely do a humanoid”, though he gave no design, partner or date. If frontier models can already orchestrate robot bodies through tools, the value in robotics may move towards whoever owns the model and the skill ecosystem.

What the demo does not show

Treat this as an interesting existence proof rather than a result.

No numbers. The page reports no success rates, trial counts, completion times or failure analysis. The authors show two tasks in one kitchen and write “more coming soon”. Skill clips are sped up, some 14×, and include “recorded reattempts”.
The idea is not new. Google’s SayCan (2022) already had a language model pick from a library of learned skills, with an 84% plan success rate on 101 tasks. The new part is the combination of a much stronger model, a humanoid body, and a self-built digital twin used as memory.
Latency and cost. The authors themselves list pauses between skills while Astra reasons, extra setup time and API cost for Real2Sim, and finger servos that overheat during extended operation. A cloud round trip in the control loop also means the robot depends on the network.
Five skills is a small vocabulary. Tidying a kitchen with pick, place and drawers is very different from folding laundry or handling deformable and fragile objects, which is exactly where end-to-end VLAs are trying to win.
Safety. The GPT-6 Astra system card covers cyber, bio and alignment risks but does not discuss physical control of robots. That is a notable gap given this month’s reports of OpenAI agents probing government websites. A model that improvises with tools in a browser can also improvise with a body.

The lesson for engineers is concrete even if the claim stays modest: as models improve, the bottleneck in robotics looks more and more like API design, state management and error handling. Those are problems software developers already know how to solve.