Neural-guided symbolic evolution: From symbolic control search to mechanism discovery in swarm robotics
Chen Wang, Cheng Zhu, Minfang Lu, Jiangyao Bai, Linbo Qiao · Engineering Applications of Artificial Intelligence · 2026
Extracting explicit governing rules from evolved strategies remains a challenge in swarm robotics. To address this issue, we propose a neural-guided symbolic evolution (NGSE) framework. Rather than treating controller design only as an optimization problem, we formulate it as an interpretable mechanism-discovery process. By modeling symbolic evolution as a sequential decision problem, NGSE uses a graph neural network to guide Monte Carlo tree search, enabling efficient identification of compact velocity controllers in a large symbolic search space. A key component is a large language model (LLM)-based interpreter that analyzes the evolved formulas after the search stage. This module relates symbolic expressions to physical mechanisms, such as virtual damping and potential-field shaping. Experiments on flocking, line-formation, and entrapping tasks show that NGSE achieves performance comparable to deep reinforcement learning baselines while retaining explicit coordination logic.