TL;DR
Stanford and Caltech researchers wired GPT-6 Astra directly into a Unitree G1 humanoid robot, skipping the traditional trained control layer, and had it autonomously tidy an unfamiliar kitchen, a concrete proof-of-concept for LLM-native robot control.
What happened
- Stanford and Caltech built HomeBody, a system connecting GPT-6 Astra directly to a Unitree G1 robot for open-ended kitchen tasks.
- No trained control layer: Astra calls straight into a modular skill library covering grasping, navigation, and drawer-opening.
- The robot first explores the room, builds a digital twin in Nvidia Isaac Sim, and logs object locations in spatial memory for out-of-view retrieval.
- Given a prompt like "clean up the kitchen," Astra plans each step and self-corrects on failures in real time.
- Code is public on GitHub; known limitations include Astra latency, overheating finger servos, and high compute costs.
Why it matters
- Removes the bottleneck: eliminating the trained control layer means any capable VLM can be swapped in, dramatically shortening the path from new model to working robot.
- Spatial memory lets the robot track objects it can no longer see, a core requirement for real household utility that prior systems struggled with.
- Earlier benchmarks confirmed Astra's improved spatial reasoning, but a separate study flagged safety risks when Astra directly controls hardware, making this a live tension.
- OpenAI has announced a robotics push, including personal-use robots, so HomeBody's architecture could inform or accelerate that roadmap.
- The kitchen domain is a deliberate stress test: unstructured, object-dense, and requiring multi-step planning, making success here a meaningful signal for commercial viability.
What to watch next
- Whether OpenAI adopts or endorses a HomeBody-style direct-control architecture in its own robotics products, which would validate the approach at scale.
- Progress on the safety and latency issues: Astra's response lag and servo overheating are the two most concrete blockers between lab demo and consumer deployment.
- Competing labs (Google DeepMind with Gemini Robotics, Figure, Physical Intelligence) responding with their own direct-VLM control demos, which would signal an industry-wide architectural shift.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.