An improved Q-learning algorithm for search and rescue UGV path planning in unknown environments

To minimise risks associated with tasks, unmanned ground vehicles are deployed in hazardous areas to search for targets. However, in non-cooperative environments, such as disaster-stricken mines, communication signals may be compromised, making it challenging to locate trapped individuals accurately. To increase the success rate of search and rescue operations, this article proposes an online path planning algorithm that integrates grey system theory with reinforcement learning. The algorithm's dual approach enables real-time systems to make decisions with greater resistance to interference. To generate reliable state predictions, it extends the grey residual error correction model to the environmental data analysis algorithm, improving the accuracy of data analysis. Additionally, the algorithm introduces a decision weighting scheme for multi-objective decision processes, allowing for flexible adjustments to the decisions made, and helps mitigate the influence of erroneous data. The algorithm also adaptively optimises the Q-Learning action selection algorithm, emphasising key parameters and reward functions to accelerate the stable convergence of the method. Finally, various search and rescue scenarios are simulated, and the proposed algorithm is compared with several existing online planning methods to demonstrate its effectiveness. The results show that this innovative search planning method can successfully navigate and capture targets within the specified search area.