A decision-making and planning algorithm integrating reachable sets with reinforcement learning (RL) was proposed to address the limitations of traditional reachable sets in effectively handling behavioral interactions between intelligent vehicles and adjacent vehicles in dynamic and uncertain environments, as well as excessive computational complexity. An RL model was incorporated into the algorithm framework to guide multi-step decision-making, clearly defining continuous macro driving behaviors over the planning horizon. Firstly, a reinforcement learning decision model was established and formulated as a Markov decision process (MDP), with state space, action space, and reward function designed. Secondly, feasible driving regions were partitioned based on driving semantics. Lateral and longitudinal behavioral predicates were introduced to segment reachable regions at each time step into finite feasible areas via a two-stage (lateral-first, then longitudinal) segmentation. Finally, the ego vehicle’s position was predicted from RL model outputs to determine optimal driving regions and form a driving corridor. The proposed algorithm’s effectiveness was validated through long-duration cyclic tests in dynamic and uncertain scenarios and comparative analysis of typical cases. Experimental results demonstrated that, compared with existing reachable set algorithms, the proposed method achieved better overall performance in enhancing driving efficiency and ensuring safety, comfort, and real-time responsiveness.