UCSD Hao AI Lab Speeds Up GPT-6 Astra and Codex Robotics Agents by About 3×
UCSD Hao AI Lab sped up the tested robotics agents by about 3× by consolidating tool-call round trips, streamlining context, and allowing model reasoning and robot movement to run in parallel. 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, and the proportion of time spent on robot movement increased from 24% to 84%. However, each action is still based on observations from about 5 s earlier, resulting in a response delay of about 5 s after the environment changes.
Why read this
AI score 63
The experiment demonstrates how merging tool calls, compressing context, and running reasoning and movement in parallel can reduce waiting time for robotics agents.
Topics
Body
Why don't benchmarks show this? In RoboDojo, the simulator waits for the model before it steps, but a real robot doesn't wait.
Our early towel-folding runs took 13 to 37 minutes. In one of the longest, the model took 78% of the wall time, and the robot moved for only a quarter of the run.
AI Trends