Inter-vehicle docking is the problem of coordinating multiple robots to actively form and maintain physical contact. It is an important capability for autonomous surface vehicles (ASVs) and is an essential part of a wide class of missions. This article considers one such mission: the emergency response and environmental protection problem of containing a floating pollutant. We propose a solution in which multiple robots autonomously navigate so as to surround the surface matter. Before doing so, the robots dock with one another to secure specialized attachments designed to ensnare the contaminant. We describe the prototypical physical robot system developed to perform this task, and we detail the system architecture, sensing and computational hardware, control system, and visual processing pipeline. While employing multiple ASVs maximizes spatial reconfigurability, it depends on the inter-robot docking capabilities being particularly reliable. But achieving robust docking is a significant technical challenge because the water continually induces external disturbances on the control system. These disturbances are non-stationary and almost impossible to predict for unknown environments. Our system relies primarily on visual servoing within a control framework in which a variety of sensors are fused. Accurate disturbance measurements are obtained through traditional sensor modeling and filtering techniques. As the environment is a priori unknown, varies from trial to trial, and has proven difficult to model, we apply a model-free reinforcement learning algorithm, SARSA(lambda), along with specialized initial conditions which ensure stable operation, and an exploration guidance approach that increases the speed of convergence. We adopt a two-loop control scheme for visual servoing to successfully make use of feature descriptors with various (and variable) computational times. We demonstrate this approach to the docking problem with autonomous ground vehicles and ASVs. The results from several situations are compared, showing that disturbance rejection coupled with SARSA(lambda) is an effective approach.