BINR-MAPF: A Bio-Inspired Neural-Reflex Architecture for Decentralized Multi-Agent Pathfinding

ICLR 2026 Conference Submission19446 Authors

Keywords: multi-agent path finding, finite state machine, neuro-reflex models, covariance matrix adaptation evolution strategy

Abstract: This paper presents a novel bio-inspired algorithmic framework for decentralized multi-agent path finding (BINR-MAPF), which integrates decentralized neuro-reflex behavioral models into the MAPF problem. Inspired by cockroach nervous system responses and group dynamics, we design a system where each agent employs reactive vector fields for goal attraction and collision avoidance. A finite state machine (FSM) governs behavior switching, enabling agents to adapt to local congestion and blockages. The system integrates centralized evolution strategies to optimize reflex parameters and role assignments. Experiments on grid-based maps demonstrate enhanced scalability, real-time performance, and reduced collision rates compared to baseline reactive and learning-based methods. This work bridges bio-neurological modeling and scalable swarm path finding under limited communication.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 19446