← Back ICRA 2026

Knowledge Synthesis in Dynamic Human-Swarm Interactions using LLMs

Boubacar Ballo, Absera Yihunie, Lilly Schwarzenbach, Hanan SALAM, Eliseo Ferrante

PDF

AI summary

Key figure (auto-extracted from paper)

Swarm agents equipped with LLMs can effectively aggregate and synthesize dynamic human-generated information, with social learning significantly boosting environmental coverage.

Swarm robotics Large Language Models Human-swarm interaction Knowledge synthesis Dynamic environments Collective exploration

Problem

Swarm robots struggle to build rich environmental descriptions due to cognitive limitations and the transient nature of dynamic human interactions. This gap hinders effective collective exploration in unpredictable settings.

Approach

Agents explore independently while using an LLM to integrate and compress incoming information from humans or peers into a bounded knowledge base. This synthesized knowledge is then shared with nearby agents to enable collective learning.

Key results

Agents achieve full environmental coverage when information is repeatedly presented
Single-pass information leads to incomplete coverage for both learning modes
Social learning significantly outperforms individual learning by enabling peer-to-peer knowledge sharing
LLM-integrated knowledge base successfully synthesizes dynamic inputs within a 250-token limit

Why it matters

This framework enables robust, adaptive information aggregation in unpredictable environments, advancing practical human-swarm collaboration for exploration and rescue missions.

Abstract

No abstract on file.

Index terms

Distributed Robot Systems Swarm Robotics Environment Monitoring and Management