CONTEXT-AWARE LEARNING FOR AUTONOMOUS ROBOTIC DEPLOYMENTS IN UNKNOWN ENVIRONMENTS

Abstract

A lot of robotic missions require a robot to plan and follow trajectories to execute designated tasks, such as coverage, search, and mapping. In this thesis, we show that through learning to estimate long-term efficiency, deep reinforcement learning can further improve robot performance in various missions in unknown environments. Specifically, we introduce a context-aware learning framework that empowers robots to form a context (i.e., the global representation) of the entire robot belief, which is used to sequence local/short-term movement decisions that optimize global/long-term objectives. The proposed framework leverages learned attention mechanisms for their powerful ability to capture dependencies in data across multiple spatial scales. Such capability enables the learned context to extract subtle features from the robot belief. We validate the proposed learning framework on multiple missions with different objectives and belief representations in unknown environments: (1) informative path planning, (2) autonomous exploration, and (3) navigation in unknown environments. We conduct experiments both in simulation and on hardware, where the learned planners show significant advantages over state-of-the-art approaches, in terms of both trajectory efficiency and computing time.