Research on DQN-LSTM-Based SLAM Path Optimization Method in Multi-modal Environments

To enhance traditional path planning algorithms’ adaptability in dynamic environments, this study proposes a Deep Q-Network (DQN)-Long Short-Term Memory (LSTM) framework combining Deep Reinforcement Learning (DRL) and LSTM networks for Simultaneous Localization and Mapping (SLAM). The model constructs an end-to-end navigation decision-making system by fusing spatiotemporal features from multi-modal sensors, i.e., LiDAR, Inertial Measurement Unit (IMU), and Global Positioning System (GPS), combined with an improved PointNet and a bidirectional LSTM temporal modeling module. A 1D convolutional network extracts geometric features from point clouds during feature encoding. At the decision layer, a refined reward function with Euclidean distance penalties and motion smoothness constraints accelerates policy optimization in dynamic settings. Evaluations show the DQN-LSTM model achieves a Mean Squared Error (MSE) of $9.64 \times 10^{-4}$ on standardized benchmarks, outperforming standalone DQN and LSTM by 38.4% and 94.3% respectively. The 90% error range tightens to $[0.9,1.7]$ meters. Ablation analysis confirms the LSTM layer’s 55.3% accuracy enhancement and the trainable PointNet’s 41.7% robustness improvement in sensor-limited conditions. This framework offers a highprecision navigation solution for autonomous systems in complex environments, including warehouse logistics.