Multimodal AI-Based Robot Autonomous Navigation and Decision Optimization in Complex Dynamic Environments

Traditional single-modal navigation pipelines suffer from illumination variance, occlusions, and poor generalization under highly dynamic scenarios. We propose a four-stage closed-loop framework- “multimodal perception, feature fusion, navigation modeling, decision optimization”-to boost the autonomous mobility of service robots. First, a Multimodal Transformer (MM-Transformer) is designed to align visual, LiDAR, and IMU features through spatio-temporal cross-attention and adaptive weighting. Second, a hybrid “geometric + semantic” dynamic map is updated at 10 Hz while a novel Dynamic-RRT* planner fuses kinodynamic and social constraints. Third, a multi-objective Proximal Policy Optimization (PPO) policy is trained to trade off localization accuracy, path efficiency, and energy cost. Extensive experiments on a TurtleBot4 platform are conducted in indoor corridors (high pedestrian density) and outdoor campus roads (mixed traffic). Compared with the baseline LOAM+EKF fusion, our method reduces Absolute Trajectory Error / Relative Pose Error by $\mathbf{2 5} \boldsymbol{\%} \boldsymbol{-} \mathbf{3 0} \boldsymbol{\%}$, improves dynamic obstacle avoidance success rate by $\mathbf{1 5} \boldsymbol{\%} \boldsymbol{-} \mathbf{2 0} \boldsymbol{\%}$, and yields 20% $\mathbf{2 5} \boldsymbol{\%}$ higher composite decision score while keeping decision latency below 100 ms. The proposed framework offers a deployable and cost-effective solution for robust robot navigation in real-world dynamic environments.