The robot’s “GPT moment”—this time, it genuinely requires GPT. Stanford’s latest project HomeBodyintegrates OpenAI’s flagship model GPT-6 Astra with the domestic humanoid robot Unitree G1enabling it to follow natural-language instructions to organize objects, discard cardboard boxes, and even locate medication outside its field of view and deliver it directly into a person’s hand—all within an unfamiliar kitchen environment. The scope of large-model-controlled robotic tasks has expanded significantly.
What is HomeBody?
HomeBody is an embodied intelligence project released by the Stanford team; its core idea can be summarized in one sentence:Let GPT invoke robot skills as toolsGPT-6 Astra handles goal understanding and step planning, then invokes skills such as navigation, grasping, and drawer opening to complete tasks step by step. Compared with Robocurve’s earlier approach—using GPT to control robotic arms for tabletop pick-and-place—this time G1 enters the kitchen, where the robot must not only decide what to grasp but also remember object locations and chain walking, grasping, and placing into a coherent sequence.Claude More notably, the team states that when deploying the system into a new kitchen,
no additional training data collection is required, nor is retraining a dedicated action policy necessary.Its generalization capability is exceptionally strong.A three-step workflow: first explore, then transfer to simulation, finally autonomously plan actions
HomeBody achieves “zero-data” deployment through a clear three-step workflow:
Step 1 Exploration and mapping
- The robot’s first instruction is “You are a kitchen robot—explore this space.” Astra guides G1 to move while recording camera footage, LiDAR scans, joint poses, and traversed positions, building a map via SLAM and saving this “spatial memory.”Step 2 Transfer to simulation (Real2Sim)
- Astra constructs a “digital kitchen” in the Isaac Sim simulation platform and feeds the SLAM-derived geometric information to the model, constraining reconstruction with real-world measurements. Thus, captured imagery corresponds not merely to “I’ve seen this drawer,” but to a precise, retrievable location.Step 3 Autonomous action planning
- When the user says, “Tidy up the kitchen—put the coffee bag on the island and discard the milk and orange juice cartons,” Astra integrates current visual input, the map, spatial memory, and prior simulation results to select which skill to invoke and which target to manipulate.The medication retrieval demo best illustrates the value of this memory: upon hearing only “Bring me my medicine,” the system locates the drawer from previously saved key frames and chains together navigation, drawer opening, and medication retrieval.
GPT handles “planning,” while the body is controlled by the skill library
A common division of labor in robotic foundation model systems is:
Task planning → Action generation → Body controlHomeBody modifies the middle layer—GPT-6 Astra (functioning as “System 2”) directly invokes the skill library; skill modules handle concrete action planning, and the underlying control system executes them, bypassing a learned VLA model entirely.The skill library contains pre-written, reusable motor skills such as Navigate, Pick, Place, Open drawer, and Pick from drawer. These skills share a unified interface: the foundation model need only specify “where to go, what to grasp, and where to place it”; the skill executes autonomously and returns the result. For grasping, for example, Astra simply points to the target object in the image and specifies left or right hand—the internal perception module localizes the object, estimates depth, computes the grasp pose, and the motion planner determines how the arm should reach while avoiding obstacles.
The skill library contains pre-written, reusable motor skills such as Navigate, Pick, Place, Open drawer, and Pick from drawer. These skills share a unified interface: the large model only needs to specify “where to go, what to pick, and where to place it,” after which the skill executes autonomously and returns the result. For example, during picking, Astra simply clicks the target object in the visual feed and specifies whether to use the left or right hand; the skill’s internal perception module then outlines the object, estimates depth, and computes the grasp pose, while the motion planner determines how the robotic arm should extend and navigate around obstacles.
What does this mean for embodied intelligence?
HomeBody’s value lies in further validating the path of “general-purpose large model + reusable skill library.” Past robots typically required task- and environment-specific policy training—costly and poorly transferable—whereas HomeBody demonstrates that a sufficiently capable multimodal large model (GPT-6 Astra) paired with a well-defined skill interface enables robots to operate zero-shot in unfamiliar environments.
Of course, it is not yet as simple as “buy a G1, log into GPT, and use it.” This remains a research project; hardware cost, safety, and stability over complex, long-horizon tasks still face significant challenges. Yet it provides a clear, reproducible blueprint for the direction of “large-model-driven general-purpose household robots.”
Frequently Asked Questions (FAQ)
What is HomeBody—and can ordinary people buy and use it?
HomeBody is an embodied intelligence research project by the Stanford team, designed to validate the feasibility of large-model-controlled robotics. It is currently a research demo, not a commercially available product.
How much does the Unitree G1 robot cost?
The Unitree G1 is a domestically developed humanoid robot, with an official starting price of approximately RMB 99,000 (subject to official confirmation). It targets developers, research institutions, and enterprises—not yet the general consumer market.
What is GPT-6 Astra?
GPT-6 Astra is OpenAI’s flagship multimodal model, featuring strong capabilities in understanding, reasoning, and tool invocation—it serves as the “brain” controlling the robot in this work. For the latest on OpenAI’s models, see our GPT-6.1 Sol benchmark and GPT-6 Sol and Luna benchmarking report.
What is embodied intelligence?
Embodied AI refers to endowing AI with a “body” so it can perceive and act within the physical world. Unlike chat-only AI, embodied AI must enable robots to see, move, and grasp—ultimately performing real-world tasks like humans do.
Is large-model-controlled robotics reliable—and when will it be usable?
The technical direction has been validated, but mass production remains distant. Current implementations are primarily research demos; hardware cost, safety, and long-horizon process stability still require breakthroughs. However, with advances in world models (e.g., NVIDIA Cosmos 3) and general-purpose large models, the deployment timeline for household robots is accelerating.
Summarize
What makes HomeBody most compelling is not merely that the robot can pick up objects—but that it proves the “general-purpose large model + reusable skill library” approach enables zero-data operation in unfamiliar environments. When GPT begins treating robotic actions as “tools” to invoke, the assistant robot that proactively tidies your kitchen or helps you locate medicine draws one step closer to reality.
Interested in the latest AI developments? See our AI Model Library and Tool Comparison Engineor continue reading:NVIDIA Nemotron 3 Ultra Evaluation · Kuaishou Keye-VL-2.0 Review · Gemini 3.8 Live Review · GLM 5.3 Evaluation.