UCSD makes robot agents 3x faster: towel folding drops from 18 to 6 minutes
k7agar · x · 2026-11-20
UC San Diego's FastRUA team published a deep dive on getting robot-use agents (RUAs) to run near real time on real hardware:
- Problem: GPT-6 Astra + Codex agents driving real robots are slow (Anthropic's Fable 5.2 takes 1 hour even with pauses cut from its puzzle-cube video), but benchmarks like RoboDojo hide this because synchronous simulators wait for the model before stepping.
- Bottleneck: most time goes to per-request overhead on every tool-call round trip, above all time-to-first-token (TTFT).
- Fixes: cut request count, keep context short, speed up tools, and pipeline thinking and motion — the robot executes actions while the model plans the next ones.
- Results: towel folding dropped from 18 min to 6 min median, a cup task from 11.1 to 3.4 min, and the robot now moves 84% of the time (up from 24%). Remaining latency is 5 seconds.
Code, robot API, agent harness and a small simulator are open-sourced on GitHub.
More from Embodied
- NUS MAGIC Lab hiring two PhD students in data-efficient robot learning — DJiafei · 2026-11-20
- Machina Labs launches Machina Rebellis, its Tacoma reimagined by AI robot factory in 2.5 weeks — bennash · 2026-11-20
- Tesla Optimus V3 spotted: production line video shows head and dexterous hand — CyberRobooo · 2026-11-20
- After her grandfather's fatal fall, she names the home robot feature that matters most — paigeinsf · 2026-11-20
- Maker Builds a Flying Probe Machine from Two Broken Ender3 Printers — kscottz · 2026-11-20
- Robots will proliferate far faster than new homes get built, argues roboticist — KyleMorgenstein · 2026-11-20