2020/06/29 by Travis Manderson, Manderson, Travis, Juan Camilo Gamboa Higuera +11 · 3 citations
Computer Science · Engineering · Environmental Science · #Coral and Marine Ecosystems Studies #FOS: Computer and information sciences #Multimodal Machine Learning Applications #Robotics (cs.RO) #Underwater Vehicles and Communication Systems
paper · pdf · doi:10.48550/arxiv.2006.16235
openalex publication_date 2020/06/29 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
We present Nav2Goal, a data-efficient and end-to-end learning method for\ngoal-conditioned visual navigation. Our technique is used to train a navigation\npolicy that enables a robot to navigate close to sparse geographic waypoints\nprovided by a user without any prior map, all while avoiding obstacles and\nchoosing paths that cover user-informed regions of interest. Our approach is\nbased on recent advances in conditional imitation learning. General-purpose,\nsafe and informative actions are demonstrated by a human expert. The learned\npolicy is subsequently extended to be goal-conditioned by training with\nhindsight relabelling, guided by the robot's relative localization system,\nwhich requires no additional manual annotation. We deployed our method on an\nunderwater vehicle in the open ocean to collect scientifically relevant data of\ncoral reefs, which allowed our robot to operate safely and autonomously, even\nat very close proximity to the coral. Our field deployments have demonstrated\nover a kilometer of autonomous visual navigation, where the robot reaches on\nthe order of 40 waypoints, while collecting scientifically relevant data. This\nis done while travelling within 0.5 m altitude from sensitive corals and\nexhibiting significant learned agility to overcome turbulent ocean conditions\nand to actively avoid collisions.\n