In this paper, we present a novel underwater autonomy architecture that combines reasoning about prior history with dynamic selection of the best set of action options for the current environment. Prior history is recorded as a set of indexed episodes, including features best describing the environment each episode occurred in, the set of actions and their parameters, and the system’s performance during the episode. Based on previous history most related to current experience, the robot dynamically selects actions and/or parameters most likely to succeed in the immediate environment; the action space is represented as a dynamic Hierarchical Task Network (HTN). We have implemented and tested the architecture in UWSim, on a simulated Blue ROV-2 with a 4 DOF manipulator in a 3D motion planning domain, where the task goal is to touch a designated underwater object with the arm’s end effector. We have shown that after just 20 episodes of learning, the robot converges on a stable global policy that maximizes success rates of object touch task. The architecture is designed to be relatively domain-independent, and is applicable to a variety of underwater tasks, such as survey/search, manipulation, active perception, etc. We are currently extending our implementation to a survey domain.
|