Reward Engineering with Large Language Models for Multi-Agent Formation Control

Abstract

This study explores LLM-guided reward shaping for multi-agent formation control with collision avoidance. LLMs iteratively generate and refine reward functions based on high-level task objectives and observed failures. Evaluation across 600 simulated and real-world episodesdemonstrates that the approach achieves target formations with 35.2% fewer training iterations and improves collision-free success rates by 28.4% compared with manually designed rewards.