Multi-agent LLM robot search: 6 agents cut search time 2x, theory bound within 3-7x of observed
peter_richtarik · x · 2026-10-22
Richtarik's team bridges theory and practice in multi-agent LLM-coordinated robot search: on a two-floor task, 18 robots guided by 6 LLM agents finish with a median search time of 595s vs 1200s for a single agent—about 2x faster.
Theory-meets-practice numbers (see chart):
- Measured constants: geometry μ=0.43–0.70, local search q(3)≈0.66, realized colors κ≈1.2–1.5 (vs κ=3 in theory)
- LLM agents capture about 2/3 of the available improvement in H=3 steps
- The bound holds in every round of every run; with κ, it lands within 3–7x of the observed gap (bound 50 vs observed 13, vs 396 with κ=3)
- Merging the best verified combination of proposals (even interfering ones) instead of one color cuts mean search time by another 29–36%
More from Embodied
- NVIDIA shows one post-training recipe turns Cosmos 3 into real-time world simulators for robots and AVs — liu_mingyu · 2026-10-22
- Unitree robot takes a car bumper hit on the street and the jokes write themselves — wavefnx · 2026-10-22
- LLM agents coordinate robot swarms: correction fixes collisions, teams search 2x faster — peter_richtarik · 2026-10-22
- Generative Bionics and AMD to Build 5,000㎡ Physical AI Data Factory in Italy — CyberRobooo · 2026-10-22
- Tesla Summon works at best 2% of the time, owner reports — geoffwolfe · 2026-10-22
- RL researcher: domain-randomized policies inherently cover the friction distribution — ChongZzZhang · 2026-10-22