AI Trends

Hao AI Lab Makes GPT-6 Astra and Codex Robotic Agents About 3× Faster at Tasks

UCSD Hao AI LabUpdated Nov 20 09:58

AI summary

Hao AI Lab made GPT-6 Astra and Codex agents run about 3× faster on test tasks by consolidating tool-call round trips, shortening context, and allowing model reasoning to run in parallel with robot movement. 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%. Each action is still based on observations from about 5 s earlier, so the response delay after an environmental change is about 5 s.

Why read this

AI score 57

By consolidating tool-call round trips, compressing context, and allowing reasoning to run in parallel with robotic-arm movement, the original article presents transferable methods for reducing latency in real-world robotic agents.

Topics

Tags

Body

Awesome work done by Shaoxiong Duan @noctuafans, Yuanbo @YuanboYang60742, Quanquan Peng @QuanquanPeng03, Junda Chen @Junda_Chen_, and Hao Zhang @haozhangml.

Source

UCSD Hao AI Lab