A Path Planning Algorithm for Indoor Fire Escape On Domestic Robots by Optimised Deep Reinforcement Learning With APF Method

ABSTRACT Robots complete tasks in dangerous and extreme environments can significantly enhance people's life safety. Given the complexity and variability of indoor fire accidents, the traditional artificial potential field (APF) method faces challenges in executing the task of path planning. As two types of deep reinforcement learning (DRL) with APF methods, in this paper we leverage an on‐value method deep Q‐learning network (DQN–APF) and an on‐policy method deep deterministic policy gradient (DDPG‐APF) to address the path planning problem. For the purpose of validation upon both the algorithm's efficiency and generalisation, we take Harbin Geographic Information Industrial Park in Heilongjiang Province of China as our research area and incorporate Microsoft HoloLens devices as APF builders for simulating potential fields. Experiments show that both optimisation algorithms can significantly enhance the capabilities on path planning in complex environments. In six random obstacles environments, the DDPG‐APF is better than the DQN–APF method, with a 14.4% higher efficiency. Furthermore, for the generalisation of the algorithm, DDPG‐APF requires less time to plan a more efficient path than DQN–APF, respectively. The experimental results indicate that both optimisation algorithms effectively enable path planning for robots in indoor fire accidents and demonstrate good generalisation abilities.