Fast Autonomous Underwater Exploration using a Hybrid Focus Model with Semantic Representation

By combining complex exteroceptive perceptions, humans are able to seamlessly explore a new environment, build up their knowledge base, and “easily” identify novel objects or imminent hazards. This work aims at employing a similar intelligent cognition and exploration system for unmanned marine robots. The proposed system semantically discriminates the visual inputs of the robot's environment by combining task-directed and stimulation-based focus to encode artificial “attention”, to enhance exploration decision making. Additionally, the hybrid focus model can use the semantic descriptors to label a target object (e.g., a certain type of coral) with metric information (e.g., geographical locations of the coral) from a range-and-bearing sensor, allowing a metric map to be generated with semantic labels of the desired types of objects. These maps can help simplify human and robot interaction by providing a clear semantic interpretation of the robots observations. The proposed system is implemented and tested on an indoor robot and with underwater images.