SARP: A Semantic-Augmented Reinforcement Planning Framework for Mobile Robot Obstacle Avoidance and Path Optimization

This paper presents SARP, a Semantic-Augmented Reinforcement Planning framework for mobile robot navigation in complex and dynamic environments. By integrating semantic embeddings from a pre-trained language encoder with reinforcement learning, SARP constructs semantically enriched state representations that combine geometric perception and contextual knowledge. A semantic-guided policy optimization mechanism aligns efficiency, safety, and task intent, while asynchronous semantic updates decouple language encoding from the high-frequency control loop. Experiments in dynamic maze, corridor crossing, and crowded field scenarios show that SARP substantially outperforms conventional reinforcement learning baselines and achieves competitive or slightly superior navigation performance compared with representative large-language-model-aware approaches, while requiring substantially lower online semantic inference overhead. Profiling on desktop and embedded NVIDIA platforms further demonstrates a favorable performance–latency trade-off for edge deployment.