Learning a robotic manipulation task often involves several actions that must be completed in the right order. This challenge becomes even more important underwater, where contact-rich tasks must be performed reliably and autonomously. We have developed a learning pipeline, currently under evaluation for publication, that enables a single AI policy to track each stage of a task and adapt its actions as the task progresses. Rather than requiring separate controllers for each step or relying on high-level controllers to simplify the action space, the method enables a single policy to learn the entire task directly at the joint level. The architecture can be applied to different manipulation scenarios. While we have started with a fixed-base, single-arm task, we aim to extend it to mobile-base, bimanual tasks. In simulated experiments, it has achieved success rates above 97%, supporting its potential for robust multi-stage underwater intervention.