Improving Multi-Robot Cooperative Navigation Through Bio-Inspired Reinforcement Learning Based Local Sensing and Control Approaches
Year: 2025
Type: preprint
Abstract: Multi-robot cooperative navigation in dynamic environments presents vital challenges in coordination, safety, and scalability that existing approaches have not effectively addressed. This paper presents a novel bio-inspired reinforcement learning (BIRL) model that enables distributed decision-making through local sensing and adaptive behavioral primitives derived from natural swarm systems. The method integrates alignment, cohesion, separation, and pheromone-based communication mechanisms within a distributed actor-critic architecture that learns adaptive policies while maintaining natural coordination patterns. The mathematical background formulates cooperative navigation as a multi-objective optimization problem encompassing collision avoidance, goal convergence, path efficiency, and energy conservation. Bio-inspired behavioral primitives operate exclusively on local data without requiring centralized coordination, while sensor fusion combines multi-modal measurements through Extended Kalman Filtering with attention mechanisms that adapt based on behavioral context. Comprehensive experimental validation demonstrates substantial performance advantages over established baseline methods. The proposed method achieves a 96.2% mission success rate (SR) with a 0.8% collision rate (CR), representing a 14.7% improvement in SR and a 75% reduction in CR compared to the best baseline method. Scalability assessment confirms graceful performance degradation across team sizes from 2 to 12 robots, maintaining 80 Hz control frequencies. Robustness evaluation demonstrates exceptional resilience to sensor noise (12.1% performance degradation under 20% noise), communication failures (81.7% SR with 30% PLR), and robot failures (91.7% formation recovery rate). Ablation studies validate the critical contributions of bio-inspired components, with complete removal reducing SR to 74.6% and requiring 77% more training episodes.
Authors: Sudhakar Sengan, Mahmoud Ali Ali, Menaga D, Sivakumar G, Aseel Smerat, Firas Tayseer Ayasrah, Umidbek Abdalov
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4414550959
Cites:
Cited by:
Related to: