Multimodal Feature Fusion for Deep Reinforcement Learning-Based Mobile Robot Navigation

초록

Robust navigation in complex and dynamic environments poses a significant challenge for reinforcement learning (RL) algorithms. Traditional RL methods often rely on single-modality inputs, such as LiDAR or camera data, which limits their perception and decision-making capabilities in real-world applications. In this paper, we propose a multimodal feature extraction network integrated with a Soft Actor-Critic (SAC) framework to address these limitations. Our method combines Bird's-Eye View (BEV) images, distance data, and target-related information to generate a unified feature representation. The network incorporates convolutional layers and Layer Normalization to effectively fuse multimodal features while mitigating gradient instability. Furthermore, we employ a Gated Recurrent Unit (GRU) to model temporal dependencies, enabling the agent to adapt to dynamic environmental changes. To evaluate the proposed approach, we conducted extensive experiments on a navigation task in a complex environment. Results demonstrate that our method significantly outperforms baseline methods in terms of training efficiency, policy robustness, and reward convergence. Notably, the inclusion of temporal modeling via GRU enhances the agent's ability to capture dynamic environmental features, leading to superior performance compared to methods without temporal modeling. These findings highlight the potential of multimodal reinforcement learning frameworks for real-world robotic applications.

키워드