Human-Guided Learning for Personalizing Robot Behaviors

Abstract

The presence of robots performing tasks in real-world environments is rapidly growing. These robots will interact with various humans with different personal preferences, highlighting the need for robots that adapt their behavior accordingly. In this thesis, we develop tools and interfaces to convey task-critical information and personalize robot behavior. First, we explore settings where humans provide demonstrations for multiple tasks. For this setting, we introduce PECAN (Personalizing Robot Behavior through a Learned Canonical Space), a learning and interface-based approach that enables users to directly select their desired style. PECAN learn a continuous canonical space from demonstrations, where each point in the space corresponds to a style consistent across each task. Our simulation experiments and user studies indicate that humans prefer using PECAN to personalize robot behavior compared to existing methods. We then examine scenarios where robots complete a task in dynamic environments. A fundamental limitation when learning from demonstrations is causal confusion due to observations containing both task-relevant and extraneous information. Because the robot does not know what aspects of its observations are important a priori, it may fail to learn the intended task. We propose RECON (Reducing Causal Confusion with Human-Placed Markers), a framework that leverages beacons (UWB trackers) attached to task-relevant objects by the human before providing demonstrations. RECON learns a compact observation embedding correlated to the beacon information, and autonomously filters out extraneous information. Our experiments indicate that RECON significantly reduces the number of demonstrations required for teaching a task to the robot.