Safe Path Planning Based on Multimodal Deep Reinforcement Learning for UGVs

Autonomous path planning of unmanned ground vehicles (UGVs) in complex and dynamic environments remains a formidable and unresolved challenge. Traditional path planning approaches usually rely heavily on pre-built maps or handcrafted perception models, which severely limit the adaptability to unknown or changing surroundings. To address these limitations, this paper presents a novel multimodal deep reinforcement learning framework named CG-D3QN-HK for safe, adaptive, and efficient path planning. Specifically, the presented framework integrates a convolutional neural network to extract spatial features from visual inputs and a gated recurrent unit to capture temporal dependencies, both embedded within a dueling double deep Q-network architecture. This integration enables effective spatio-temporal information fusion for decision-making in dynamic contexts. Moreover, a shaped reward function that incorporates directional guidance and heuristic knowledge is developed to accelerate convergence and improve path quality. Extensive simulations in diverse and unstructured environments demonstrate that the CG-D3QN-HK consistently achieves higher success rates, shorter path lengths, and greater robustness compared to conventional DRL baselines, which performs its potential for real-world autonomous navigation applications.