The authors built on grounded generative intelligence for multi-robot task planning and addressed three failure modes of LLM orchestrators in continuous missions: unbounded prompt growth, degraded reasoning with many agents and tasks, and loss of vigilance after covering the area once.
LLM orchestrator persistently coordinates robot teams in dynamic disasters
Three mechanisms—Markovian state summary, partial delegation, and staleness-triggered re-prompting—let a single LLM manage a heterogeneous robot swarm indefinitely in a spreading wildfire search-and-rescue simulation.
Top University
Maria G. Mendoza · Victoria M. Tuck · Zachary Ravichandran · Shankar Sastry · George J. Pappas
University of California, Berkeley · University of Pennsylvania
Research Digest··3 min read
Thread:Agent Harness Optimization
Mendoza et al.
Why this paper
From University of California, Berkeley and University of Pennsylvania · Part of Agent Harness Optimization, now 135 papers
In one line
A framework using Markovian summaries, partial delegation, and staleness-aware reprompting rescues 98% of civilians in continuous search-and-rescue with fewer tokens.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§