Adaptive Navigation for Robots Based on Object Relationship Reasoning and Reinforcement Learning

Abstract

This study focuses on the deep integration of object relationship reasoning and reinforcement learning, proposing an adaptive navigation framework for robots. The core research contents include designing a dynamic object relationship graph model, quantifying relationship strength through spatial, motion, and functional features to achieve real-time modeling and updating of environmental relationships, constructing a reinforcement learning system that integrates relationship features, optimizing the state space and two-layer reward mechanism to improve the semantic rationality of decision-making, and building a closed-loop adaptive mechanism of "perception-reasoning-learning-decision-making" to dynamically adjust the learning rate and exploration rate, ensuring adaptability to complex environments. Experimental results show that the framework achieves navigation success rates of 96.7% and 92.5% in indoor dynamic scenes and outdoor semi-structured scenes, respectively, with collision rates of only 1.7% and 3.3%. Compared with traditional algorithms such as PPO and DQN+SLAM, the success rate improved by 8.3%-13.4%, the collision rate reduced by 4.9%-9.1%, and the average path length shortened by 3.7%-14.4%. Despite sensor noise interference and high-density obstacle environments, the framework maintains good robustness, with a stable decision response time of 32.6ms, meeting real-time navigation requirements and providing an effective solution for robot navigation in highly dynamic and interactive scenarios.