This paper presents a control architecture for an autonomous underwater vehicle (AUV) named the Component Oriented Layer-based Architecture for Autonomy (COLA2). The proposal implements a component-oriented layer-based control architecture structured in three layers: the reactive layer, the execution layer, and the mission layer. Concerning the reactive layer, to improve the vehicle primitives' adaptability to unknown changing environments, reinforcement learning (RL) techniques have been programmed. Starting from a learned-in-simulation policy, the RL-based primitive cableTracking has been trained to follow an underwater cable in a real experiment inside a water tank using the Ictineu AUV. The execution layer implements a discrete event system (DES) based on Petri nets (PNs). PNs have been used to safely model the primitives' execution flow by means of Petri net building block (PNBBs) that have been designed according to some reachability properties showing that it is possible to compose them preserving these qualities. The mission layer describes the mission phases using a high-level mission control language (MCL), which is automatically compiled into a PN. The MCL presents agreeable properties of simplicity and structured programming. MCL can be used to describe offline imperative missions or to describe planning operators, in charge of solving a particular phase of a mission. If planning operators are defined, an onboard planner will be able to sequence them to achieve the proposed goals. The whole architecture has been validated in a cable tracking mission divided in two main phases. First, the cableTracking primitive of the reactive layer has been trained to follow a cable in a water tank with the Ictineu AUV, one of the research platforms available in the Computer Vision and Robotics Group (VICOROB), University of Girona, Girona, Spain. Second, the whole architecture has been proved in a realistic simulation of a whole cable tracking mission