Hao AI Lab Lets Robots Think While Moving, Boosting Task Speed by About 3x in Real-World Tests
Hao AI Lab pipelined the reasoning of GPT-6 Astra + Codex agents with robot motion, making test tasks run about 3x faster: the cup task was shortened from 11.1 min to 3.4 min, while towel folding was reduced from about 18 min to about 6 min. The agent writes actions as Python functions, which the robot executes sequentially in the background while returning results; meanwhile, the model plans or rewrites subsequent steps. The team also consolidated tool-call round trips and converted older images into cropped text notes. Each action is still based on observations from about 5 s earlier, so the response time after an environmental change is approximately 5 s.
Why read this
AI score 63
By running model planning and robot actions in parallel and consolidating tool-call round trips, the team offers a transferable engineering approach for reducing real-world robot task completion times.
Topics
Body
Then we pipelined thinking and motion: the robot moves while the model thinks.
The agent writes actions as Python functions, and the robot runs them in order in the background and reports back as each action ends, while the model plans the next steps or rewrites the plan. https://t.co/QwdKU1aLLZ
AI Trends