researchOpenAI

Stanford, Caltech Researchers Wire GPT-6 Astra Directly Into a Robot to Clean an Unfamiliar Kitchen

TL;DR

Researchers built HomeBody, a system that connects GPT-6 Astra directly to a Unitree G1 robot's skill library, letting it explore, map, and tidy an unfamiliar kitchen without a trained control layer in between. The team reports latency, overheating servos, and compute cost as current limitations.

2 min read
0

Researchers from Stanford and Caltech have built HomeBody, a system that connects OpenAI's GPT-6 Astra directly to a Unitree G1 humanoid robot, enabling it to autonomously explore an unfamiliar kitchen, tidy it up, and retrieve items from drawers without a task-specific trained control layer between the language model and the robot's actuators.

How HomeBody Works

Most robot-control stacks place a trained intermediary layer between a vision-language model (VLM) and the robot's motor commands. HomeBody removes that layer. Instead, the VLM — in this case GPT-6 Astra — calls directly into an extensible skill library that handles primitives like grasping, navigating, and opening drawers.

According to the researchers, the robot first explores the room, building a digital twin of the space in Nvidia's Isaac Sim simulation environment. It logs object identities and locations into a spatial memory system, which allows it to locate items even after they leave the robot's camera field of view.

When given a high-level instruction such as "clean up the kitchen," GPT-6 Astra plans each step of the task and self-corrects when errors occur, according to the team. The code for HomeBody has been published on GitHub.

Limitations Reported

The researchers identified several practical constraints in the current setup:

  • Latency: GPT-6 Astra's response time introduces delays in the robot's action loop.
  • Hardware stress: The Unitree G1's finger servos reportedly overheated during extended operation.
  • Compute cost: Running a large VLM directly in the control loop is described as computationally expensive.

These issues echo findings from earlier, separate benchmarks: one showed GPT-6 Astra has improved spatial reasoning compared to prior models, while another flagged safety concerns when Astra is given direct control over a physical robot.

Context: OpenAI's Robotics Push

OpenAI has previously stated it intends to re-enter robotics, including consumer/personal-use applications, after largely stepping away from hardware robotics research years earlier. HomeBody is an independent academic project, not an OpenAI product, but it serves as one of several public demonstrations that GPT-6 Astra — OpenAI's vision-language model — can be adapted for embodied, physical-world tasks beyond chat or coding.

No benchmark scores, success rates, or quantitative task-completion metrics were disclosed in the available material. Pricing and compute costs for running Astra in this configuration were not specified.

What This Means

HomeBody is a research prototype, not a deployable product — the overheating servos and latency issues alone rule out any near-term commercial use. Its significance is architectural: bypassing a trained control layer and letting a general-purpose VLM plan and self-correct directly against a skill library suggests that robot autonomy may increasingly ride on foundation model improvements rather than task-specific training pipelines. That's a meaningful shift if it holds up under more rigorous testing, since it would mean robotics progress tracks language-model progress more directly. But the flagged safety concerns around VLMs controlling physical robots deserve more scrutiny before this pattern scales beyond lab demonstrations. For now, treat the kitchen-cleaning result as a proof of concept, not a product roadmap.

Related Articles

benchmark

GPT-6.2 Astra scores 41.2% on XDOF's Ember-0 robot benchmark, up from 19.6% for GPT-6 Astra

Robotics infrastructure company XDOF reports that GPT-6.2 Astra reached a 41.2% average success rate across the 16 tasks of its Ember-0 benchmark, versus 19.6% for GPT-6 Astra. It is the first model to complete rope knot-tying and T-shirt unfolding on the benchmark. Average prediction latency fell from 15 seconds to 10.1 seconds.

changelog

OpenAI says GPT-6.2 Astra safeguards cut reasoning-extraction attack success to 0.2%

OpenAI announced on November 20, 2026 stronger defenses against adversarial distillation. It reports that adaptive attacks extracting protected reasoning from an earlier request succeed 0.2% of the time on GPT-6.2 Astra with updated safeguards, versus 54.0% on GPT-6 Astra with original safeguards.

product update

OpenAI tests ChatGPT dots with a small group of Plus users, rolls out to all Business Standard

OpenAI has started testing ChatGPT dots with a small number of ChatGPT Plus subscribers and is rolling it out to all ChatGPT Business Standard users. Dots launched in September for Pro, Business Premium and Enterprise plans. Usage allowances for the new tiers have not yet been clarified.

research

OpenAI's Pachocki proposes perplexity as a shared metric for industry safety limits, publishes internal figures

OpenAI Chief Scientist Jakub Pachocki argues that perplexity (cross-entropy loss) should be a shared metric for comparing model developers and informing safety standards. OpenAI is publishing internal perplexity measurements, citing log-perplexity of ~0.99 for GPT-3.5 and ~0.91 for GPT-4.

Comments

Loading...