# HomeBody: robot AI research model

HomeBody is a Stanford The Movement Lab + Caltech research system (Huh, Gu, Truong; advisors Karen Liu and Guy Tevet; Sep 2026) that puts a frontier VLM (demo: GPT Astra) in charge of a [Unitree G1](/models/unitree-g1.md) with persistent spatial memory and a composable skill library, without a learned [VLA](/glossary/vision-language-action-model.md) between high-level reasoning and motor skills. After exploring an unseen kitchen and building an [Isaac Sim](/brains/isaac-sim.md) digital twin, the system chains navigate / pick / place / open-drawer skills for long-horizon tasks. Project-stated limits: Real2Sim cost, VLM latency, finger-servo heat, RTX 4090 local compute. Academic research demo, not a commercial product. No GPT Astra model page created (LLM product held per prior pattern).

- **Slug:** homebody
- **Type:** Research model
- **Maturity (DEPLOY ladder):** Research
- **Open source:** no

## Architecture

Frontier VLM (System 2) directly orchestrates a composable skill library (navigate, pick, place, open drawer, pick from drawer) with Super Odometry / ICP spatial memory and Real2Sim twin in NVIDIA Isaac Sim; skills run locally (RTX 4090 laptop) with remote VLM tool calls; AMO lower-body control; no trained VLA in the middle


## Key facts

- **System (project-stated):** Project-stated (project-stated): HomeBody (Huh / Gu / Truong; advisors Liu and Tevet; Caltech + [Stanford](/locations/stanford.md), Sep 2026) equips a frontier VLM with persistent spatial memory and a composable humanoid skill library (navigate, pick, place, open drawer, pick from drawer) so the VLM can orchestrate long-horizon [loco-manipulation](/glossary/loco-manipulation.md) without a learned [VLA](/glossary/vision-language-action-model.md) in the middle. Demo uses a [Unitree G1](/models/unitree-g1.md) in a previously unseen kitchen. Academic system, not a commercial product
- **Explore / Real2Sim / skill chaining (project-stated):** Project-stated (project-stated): the humanoid explores an unfamiliar kitchen (iPhone video, D435i, [LiDAR SLAM](/glossary/slam.md), joint poses, Astra-chosen waypoints), builds a digital twin in [NVIDIA Isaac Sim](/brains/isaac-sim.md) from that exploration data, then chains skills for tasks such as kitchen tidy and medicine retrieval from an underspecified request. Stated limits include Real2Sim setup/API cost, GPT Astra reasoning latency between skills, finger-servo overheating on extended runs, and local RTX 4090 laptop compute for the skill stack. Academic demo
- **HomeBody architecture (primary Stanford TML):** Stanford The Movement Lab's Sept 2026 HomeBody page describes a humanoid system that equips frontier VLMs with persistent spatial memory and a composable skill library (navigate, pick, place, open drawer) so System 2 can orchestrate long-horizon [loco-manipulation](/glossary/loco-manipulation.md) without an intervening learned [VLA](/glossary/vision-language-action-model.md). Demo platform noted is a [Unitree G1](/models/unitree-g1.md) guided by GPT Astra in an unseen kitchen ([Stanford](/locations/stanford.md) TML)
- **Real2Sim and compute note (primary Stanford TML):** The same page says HomeBody uses Astra as a Real2Sim agent to build an [Isaac Sim](/brains/isaac-sim.md) digital twin from exploration data (iPhone video, D435i, [LiDAR SLAM](/glossary/slam.md), joint poses), and that current local skills run on a single Razer Blade laptop with an RTX 4090 while Astra runs remotely. Treat as research prototype claims ([Stanford](/locations/stanford.md) TML)
- **Architecture and demo platform (primary Stanford TML):** Stanford The Movement Lab / Caltech HomeBody page (Sep 2026) describes a humanoid autonomy system that lets a frontier VLM (demo: GPT Astra) directly orchestrate a composable skill library with persistent spatial memory on a [Unitree G1](/models/unitree-g1.md), replacing a learned [VLA](/glossary/vision-language-action-model.md) between System 2 reasoning and System 0 control. Kitchen demos include multi-object cleanup and out-of-view medicine retrieval without environment-specific policy learning (authors' stated claims) ([Stanford](/locations/stanford.md) TML)
- **Real2Sim and compute (primary Stanford TML):** The same page says HomeBody uses Astra as a Real2Sim agent to build an [Isaac Sim](/brains/isaac-sim.md) digital twin from G1 exploration data (iPhone video, D435i, [LiDAR](/glossary/lidar.md)/[SLAM](/glossary/slam.md), joint poses), with current skills running on a single Razer Blade laptop with an RTX 4090 while GPT Astra runs remotely. Treat hardware and pipeline details as authors-stated ([Stanford](/locations/stanford.md) TML)
- **Official project scope:** Stanford The Movement Lab describes HomeBody as a research system that lets a frontier vision-language model orchestrate a composable humanoid skill library on a [Unitree G1](/models/unitree-g1.md) in an unfamiliar kitchen (HomeBody).
- **Research prototype status:** The [Stanford](/locations/stanford.md) project page presents HomeBody as an academic humanoid-autonomy demonstration, not a commercial robot product (HomeBody).


## Sources (2)

1. **HomeBody: A Humanoid That Explores, Remembers, and Acts on Its Own** · https://tml.stanford.edu/homebody/ · 2026-09-01
2. **Secondary summary: Stanford + Caltech HomeBody puts GPT Astra in charge of Unitree G1** · https://x.com/carlosearias/status/2104218349123031470 · 2026-09-27


## Common questions

### What is HomeBody?

HomeBody is a Stanford The Movement Lab + Caltech research system (Huh, Gu, Truong; advisors Karen Liu and Guy Tevet; Sep 2026) that puts a frontier VLM (demo: GPT Astra) in charge of a [Unitree G1](/models/unitree-g1.md) with persistent spatial memory and a composable skill library, without a learned [VLA](/glossary/vision-language-action-model.md) between high-level reasoning and motor skills. After exploring an unseen kitchen and building an [Isaac Sim](/brains/isaac-sim.md) digital twin, the system chains navigate / pick / place / open-drawer skills for long-horizon tasks. Project-stated limits: Real2Sim cost, VLM latency, finger-servo heat, RTX 4090 local compute. Academic research demo, not a commercial product. No GPT Astra model page created (LLM product held per prior pattern).

### Is HomeBody open source?

No. HomeBody is recorded as proprietary on the DEPLOY registry. Model weights and source are not publicly available.

### What type of AI is HomeBody?

HomeBody is a research model, built on a Frontier VLM (System 2) directly orchestrates a composable skill library (navigate, pick, place, open drawer, pick from drawer) with Super Odometry / ICP spatial memory and Real2Sim twin in [NVIDIA Isaac Sim](/brains/isaac-sim.md); skills run locally (RTX 4090 laptop) with remote VLM tool calls; AMO lower-body control; no trained [VLA](/glossary/vision-language-action-model.md) in the middle architecture on the DEPLOY registry.

### What is HomeBody's maturity stage?

HomeBody is at the research stage on the DEPLOY maturity ladder. Research stage means active development without commercial deployments on file.

### Which robots run on HomeBody?

No robot models on the DEPLOY registry are recorded as running HomeBody. DEPLOY wires brain-to-model connections only when the wiring is verifiable from primary sources; absence may reflect pre-deployment or unverified manufacturer claims.


_API: GET /v1/brains/afb31f1b-a60f-4f94-8f0b-67d0d1b84f3e · canonical URL: /brains/homebody_