This paper addresses the trade-off between control performance and safety in reinforcement learning (RL)-based path planning by proposing RL+adpCBF, a safe RL algorithm augmented with control barrier functions (CBFs). First, within the framework of constrained Markov decision processes (CMDPs), CBFs are incorporated to provide formal safety guarantees, and an end-to-end safe RL framework is constructed based on twin delayed deep deterministic policy gradient (TD3). Second, we design the CBFNet architecture, which introduces CBF-based safety constraints under the constrained RL paradigm and achieves adaptive safety-performance balancing via the Lagrangian multiplier method. Meanwhile, the CBF constraints derived from prior dynamics and safety sets are formulated as a differentiable optimization layer, embedding explicit convex safety constraints into the policy network and enabling end-to-end gradient backpropagation. Finally, the proposed method is validated in simulation. Experimental results demonstrate that the RL+adpCBF effectively avoids unsafe actions while maintaining high control performance, responds rapidly to environmental changes with real-time policy adjustments, and significantly improves the safety and operational efficiency of mobile robots.