presentofai

GPT-6 Astra Directly Controls Humanoid Robot in Kitchen Task

Published on

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, signaling that frontier language models may soon replace purpose-built robot controllers.

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 into a modular skill library (grasping, navigation, drawer-opening) without an intermediary policy model.
  • The robot first explores the room, builds a digital twin in Nvidia Isaac Sim, and stores 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 when errors occur.
  • Code is public on GitHub, lowering the barrier for replication and extension.

Why it matters

  • Eliminates the control-layer bottleneck: swapping in a better VLM instantly upgrades the robot, no retraining required.
  • Spatial memory plus digital twin lets the robot track objects it can no longer see, a key gap in prior household robot demos.
  • Astra's improved spatial reasoning (confirmed in earlier benchmarks) is now load-bearing in a physical manipulation context, not just a benchmark score.
  • OpenAI has announced a return to robotics, including personal-use products, making HomeBody a proof-of-concept that validates that strategic direction.
  • Identified safety concerns from a separate benchmark when Astra controls a robot remain unresolved, adding regulatory and liability weight to rapid deployment.

What to watch next

  • Whether OpenAI's robotics push adopts a HomeBody-style direct-control architecture or reintroduces a trained intermediary layer for safety.
  • Progress on the three named limitations: Astra inference latency, overheating finger servos, and compute cost, as each is a hard commercial barrier.
  • Independent safety audits of direct VLM-to-robot control, given the flagged risks; a serious incident or a formal safety framework would set the trajectory for the field.

Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.