AI Trends

Hao AI Lab Speeds Up GPT-6 Astra + Codex Robotic Agent Tasks by About 3×

UCSD Hao AI LabUpdated Nov 20 09:59

AI summary

By consolidating tool-call round trips, shortening context, and allowing the model to think while the robot moves, Hao AI Lab made the GPT-6 Astra + Codex agent about 3x faster on test tasks. The cup task was reduced from 11.1 min to 3.4 min, while towel folding fell from about 18 min to about 6 min, and the share of time spent on robot movement rose from 24% to 84%. However, each action is still based on visual information from about 5 s earlier, resulting in a response delay of about 5 s after environmental changes. The blog is at https://t.co/BIaiZ2P5ch, and the code is at https://t.co/MvWxzfMXiK.

Why read this

SelectedAI score 68

The original article breaks down engineering methods for consolidating tool round trips, compressing context, and planning actions in parallel, which can be applied to reduce wait times for real-world robotic agents.

Topics

Tags

Body

However, hiding the model's time behind motion doesn't make the robot react fast. Each action is still based on what the robot saw about 5 s earlier, so it needs about 5 s to react when something changes.

We also treated Codex and Astra as a black box. With open models served for robots, the server and the harness can be designed together, leaving much more room to cut model time and be even faster.

Blog: https://t.co/BIaiZ2P5ch Code: https://t.co/MvWxzfMXiK

Source

UCSD Hao AI Lab