AI Trends

Hao AI Lab Speeds Up GPT-6 Astra Robotic Agents by Consolidating Tool-Call Round Trips

UCSD Hao AI LabUpdated Nov 20 10:00

AI summary

Hao AI Lab consolidated tool-call round trips and enabled robot motion to run in parallel with model reasoning, making the GPT-6 Astra + Codex agent run approximately 3x faster on test tasks. The cup task was reduced from 11.1 min to 3.4 min, towel folding from ~18 min to ~6 min, and the proportion of time spent on robot motion increased from 24% to 84%. The team also made each action return camera imagery and robot state simultaneously, and compressed older images into text notes to shorten the context, although actions were still based on observations from about 5 s earlier.

Why read this

AI score 59

The material presents optimization approaches involving consolidated tool calls, parallel planning and motion, and context compression, which can be applied to latency management for real-world robotic agents.

Topics

Tags

Body

For Astra at medium reasoning effort, most of it is per-request overhead, paid on every tool-call round trip.

So we merged round trips: - each action comes back with the camera views, robot state and the agent's own registered views, so looking costs no extra request - code runs as soon as it's written; the agent checks its plan while the arm moves, instead of spending an extra turn - we keep context short, pruning old images into text notes

Source

UCSD Hao AI Lab