Hold the Spot: Evolution of Generalized Station Keeping for an Aquatic Robot

In this paper, we present a strategy to evolve neurocontrollers in aquatic robots capable of generalized station keeping, that is, maintaining a position in the presence of various water flows. Evolved behaviors exhibit a variety of complex fin/flipper movements that enable the robot to react and move against changing flows. Moreover, results indicate that some sensor modalities are beneficial when the robot is placed in novel environments, though little used during the evolutionary process. Introduction. Increasingly, aquatic robotic systems are being deployed to assist humans in challenging tasks (Tan et al., 2006). Although many systems rely on control from a human, autonomy allows robots to act independently in hostile or remote environments. Station keeping, also known as station holding, involves maintaining a position against external fluid flows, and is exhibited by many biological fish (Arnold, 1974). It is also of interest for aquatic robot sensor platforms that need to remain stationary while gathering data. In such cases, the autonomous control system needs to be able to respond to changing flows. We previously examined the evolution of station keeping behaviors for individuals facing a single flow during their evaluation and evolutionary periods (Moore et al., 2013). Although successful in this task, individuals failed to maintain station when facing novel flows (i.e., flows that were not encountered during the evolutionary process). In this work, we focus on the evolution of generalized station keeping behaviors capable of handling multiple distinct flows. We investigate approaches to this problem through two separate experimental setups that utilize multiple flows during the evolutionary fitness evaluation. Evolved individuals are then evaluated in both previously seen and novel flows. Results indicate that evolved individuals are able to hold station against flows encountered within and outside of the evolutionary process, exhibiting some generalized behaviors. Furthermore, additional sensor modalities increase an individual’s ability to generalize to new flow conditions, even though they do not appear to be beneficial in environments encountered during evolution. This work provides an approach to developing a generalized control strategy for dynamic environmental conditions, and provides insight into the impact of sensory information for evolved neural controllers. Methods. The simulated robot in this study emulates the form and function of a physical device, seen in Figure 1a. Pectoral flippers are capable of continuous 360◦ range of motion, while the caudal fin is limited to a ± 90◦ symmetric range of motion. Sensory information includes inertial data (i.e., linear and angular acceleration at the robot’s center of mass) and the previous update’s motor commands. Information about the actual state of the motors is not used as most small continuous rotation servo motors do not provide this data. In two of the treatments, we include additional sensory information including the robot’s pitch, roll, and yaw, along with the xyz flow information.