Autonomous Firefighting Robots Using Reinforcement Learning

Abstract

Autonomous firefighting robots represent a transformative approach to mitigating risk and enhancing effectiveness during fire‐fighting operations in hazardous environments. This study presents a comprehensive investigation into the design, development, and evaluation of a mobile firefighting robot driven by deep reinforcement learning (DRL). By integrating advanced perception modules—comprising LiDAR‐based simultaneous localization and mapping (SLAM) for precise environment mapping and a fused RGB‑thermal imaging pipeline for robust fire detection—with a suppression subsystem featuring a water‑mist nozzle, the robot is equipped to autonomously navigate cluttered indoor spaces, identify fire sources, and deploy extinguishing actions. We employ a Deep Q‑Network (DQN) augmented with prioritized experience replay to learn optimal navigation and suppression policies in a simulation environment reflecting realistic fire dynamics, including dynamic obstacles and variable fire intensities. Training is conducted over 300,000 time steps, with the reward structure carefully shaped to balance exploration, navigation efficiency, hazard approach, and collision avoidance. The resulting policy achieves a navigation success rate of 92% and extinguishing success of 88% across fifty test episodes, yielding a 35% reduction in average time‑to‑extinguish compared to a heuristic baseline employing A* planning and threshold‑based thermal detection. Statistical analysis via two‑sample t‑tests confirms the significance of performance gains (p < 0.001), and qualitative failure‐case examination highlights areas for improvement in smoke occlusion handling.