UCSD Hao AI Lab Speeds Up Robotic Agent Tasks by About 3× Through Parallel Reasoning and Motion
UCSD Hao AI Lab has brought a GPT-6 Astra + Codex agent closer to real-time robot control, cutting the cup task from 11.1 min to 3.4 min and towel folding from about 18 min to about 6 min. By consolidating tool-call round trips, compressing context, and allowing the robot to keep moving while the model reasons, the team increased the share of time spent on robot motion from 24% to 84%. Each action is still based on an observation from about 5 s earlier, so reacting to environmental changes also takes about 5 s.
Why read this
SelectedAI score 63
This optimization consolidates tool-call round trips and allows the model to reason while the robotic arm moves, providing a transferable method for reducing latency in real-world robotic agents.
Topics
Body
Then something harder: folding a crumpled towel (video at 8x).
In both runs, most of the model's time was hidden behind the robot's motion, and the robot stood still only 16% and 22% of the time. https://t.co/UpP0iW2oQa
AI Trends