In this work, we have discussed the necessities of collision avoidance systems in today’s world, a domain increasingly driven by automation that has taken the world by storm. For a system to be good at collision avoidance, the most important part is real-time decision-making so that it can avoid not only the stationary but also moving objects effectively. For this, we are taking a new reinforcement learning (RL) approach to improve path selection and collision avoidance. Our method makes use of an agent based RL model that integrates real-time obstacle detection and continuous environmental feedback, allowing drones to dynamically modify their navigation strategy in response to shifting environmental conditions. Proximal Policy Optimization (PPO) in conjunction with transformer-based architecture for coordinate-based data processing is used to train this framework in a simulated environment. PPO offers a strong RL backbone that adjusts to intricate, multi-dimensional environments and is selected for its steady convergence in continuous control applications. As we are using transformers, it helped to enhance the system’s capacity for obstacle identification and classification because of its ability to analyse and understand complicated, high-dimensional visual data. After doing our research, we came to the conclusion that our RL-based model outperforms conventional path planning algorithms by a great margin. We are providing a multi-functional RL framework for autonomous drone navigation, also this work provides a more safer, more resilient, and adaptive aerial systems for reliable drone operations in challenging urban environments. This study provides its uses in a larger field of autonomous systems, especially in complicated and densely inhabited environments.