Researchers led by Gengze Zhou introduced NavGPT-3, a runtime that links a large language model with a low‑latency action policy for robots. The harness runs reasoning, acting and monitoring as separate threads, letting the robot interrupt and switch tasks quickly. Their 8‑billion‑parameter action policy, trained on 19.28 million examples, already hit 74.5% success on the R2R‑CE benchmark. With the full system, NavGPT-3 improved to 81.5% on that test and matched human performance on RxR‑CE, completing routes in about 1 minute 22 seconds per episode.
Why it matters
It shows that robots can combine high‑level language reasoning with fast low‑level control, bringing autonomous navigation closer to human speed and reliability.
