The capacity of detecting and avoiding collisions among ships without human intervention is a key requirement for future cyber-ships. This work investigates a deep reinforcement learning (DRL)-based framework for multi-vessel collision avoidance, which enables autonomous collision avoidance and path planning for ships in complex maritime scenarios. The proposed method addresses key challenges such as dynamic encounter geometries, limited inter-ship communication, and adherence to COLREGs by modeling the decision process under partial observability. Three navigation strategies–velocity obstacle, deep Q-network, and recurrent LSTM-DQN–are evaluated within a unified simulation environment. Both a hexagon-based ship domain and a real-time distance-based safety metric are incorporated to assess collision risk and ensure rule compliance. The framework is validated on multiple Imazu-inspired test cases, covering diverse encounter types including head-on, overtaking, and starboard crossing. Experimental results show that the LSTM-DQN model consistently achieves safer trajectories, lower minimum distances, and higher compliance rates compared to baseline methods. These findings demonstrate the potential of our DRL-driven policies to support reliable, scalable, and regulation-aware decision-making for autonomous surface ships in a realistic environment.