Autonomous ocean-exploring vehicles promise to transform the rate at which we can explore ocean environments. Such robots, however, face the challenge of navigating through unknown ocean currents and seeking out areas of interest without prior knowledge of their surrounding environment. Inspired by the ability of aquatic animals to navigate via flow sensing, we constructed a palm-sized robotic swimmer platform to test flow-based navigation strategies in a 6 by 6 by 16-foot water tank. The robot is equipped with distributed pressure sensors to mimic the function of canal neuromasts found in the lateral lines of fishes. A deep reinforcement learning algorithm runs onboard a high-speed microcontroller, which trains a neural network in real-time to control the robot’s thrusters based on sensor inputs. As an analogy for tracking hydrothermal vent plumes in the ocean, the robot is tasked with locating a turbulent jet flow using data from the pressure sensors and insights from the flow physics of the turbulent plume.

Primary Presenter: Peter Gunnarson, Caltech (pjgunnar@caltech.edu)

Authors:

Peter Gunnarson, Caltech  (pjgunnar@caltech.edu)

John Dabiri, Caltech  (jodabiri@caltech.edu)