Space exploration is often too hazardous for humans to perform, relying on autonomous rovers and probes to safely operate. Sea navigation presents a similar challenge within a non-modelled open environment, with autonomous surface vehicles (ASVs) being developed to operate without direct human control. Before being integrated into actual traffic, ASVs must follow the International Regulations for the Prevention of Collisions at Sea (COLREGs) However, automating these rules is challenging since they have multiple interpretations depending on the situation. A COLREG-compliant framework is proposed where the ASV predicts the actions of nearby vessels. Most of the time, it will abide by the rules, except in extreme situations. Then, the ship must find a safe strategy to avoid collisions and should be able to depart from COLREGs if necessary. This article shows how Deep Reinforcement Learning can be used to achieve this task, specifically the Proximal Policy Optimization. This method has been chosen because it is efficient at creating a policy solving the dual problem of path planning and collision avoidance while following rules. This framework has potential applications beyond maritime applications, such as spacecraft collision avoidance in congested orbits, highlighting the broader implications of this work in scenarios requiring protocol adherence in hazardous environments.