Citation

Chi, Jie Tan and Munjer, M.A. and Pan, Wei Zheng and Hayashi, Eiji and Way, Soong Lim LLM-Supervised Adaptive Mutation Control for Evolutionary Behavior Tree Optimization in Multi-Objective Coastal Cleanup Robotics. Journal of Robotics, Networking and Artificial Life, 11 (4). pp. 291-298. ISSN 2352-6386

Text 11404.pdf - Published Version Restricted to Repository staff only Download (844kB)

Abstract

This study investigates the use of large language models (LLMs) as adaptive supervisory controllers for evolutionary optimization of behavior trees in multi-robot systems. Unlike conventional genetic programming approaches that rely on fixed mutation and crossover parameters, the proposed framework integrates an LLM as a high-level reasoning module that dynamically adjusts evolutionary parameters based on observed population dynamics during the search process. This enables the system to maintain a balance between exploration and exploitation without requiring manual tuning of evolutionary settings. A comparative study is conducted between the proposed LLM-guided evolutionary framework and a conventional genetic algorithm with static parameters in a coastal cleanup behavior tree optimization task. The results demonstrate that incorporating LLM-based supervision leads to consistently improved optimization performance and more effective search behavior. In particular, the proposed method achieves higher-quality solutions and faster convergence to high-fitness regions compared to the baseline approach, while maintaining greater behavioral diversity throughout the evolutionary process. Further analysis shows that LLM-guided parameter adaptation enhances the generation of diverse and specialized behavioral strategies, particularly in complex multi-objective scenarios involving task execution, resource management, and environmental interaction. These findings suggest that integrating language-model-based reasoning into evolutionary search can improve both the efficiency and robustness of behavior tree optimization, offering a promising direction for adaptive robotic control system design. © 2026 ALife Robotics Corporation Ltd. All rights reserved.

Item Type:	Article
Uncontrolled Keywords:	Behavior tree, Field robotics, Large language models (LLMs), Multi-objective optimization
Subjects:	Q Science > QA Mathematics > QA71-90 Instruments and machines
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Suzilawati Abu Samah
Date Deposited:	28 Oct 2026 07:02
Last Modified:	28 Oct 2026 07:02
URII:	http://shdl.mmu.edu.my/id/eprint/16942

Downloads

Edit (login required)