Teaching robots about human environments: Leveraging human interaction to efficiently learn and use multisensory object affordances

Abstract

The real world is complex, unstructured, and contains high levels of uncertainty. Although past work shows that robots can successfully operate in situations where a single skill is needed, they will need a framework that enables them to reason and learn continuously so that they can operate effectively in human-centric environments. One framework that allows robots to aggregate a library of skills is to model the world using affordances. In this thesis, we choose to model affordances as the relationship between a robot's actions on its environment and the effects of those actions. By modeling the world with affordances, robots can reason about what actions they need to take to achieve a goal. This thesis provides a framework that allows robots to learn affordance models through interaction and human guidance. Within the scope of robot affordance learning, there has been a large focus on learning visual skill representations due to the difficulty of getting robots to interact with the environment. Furthermore, utilizing different modalities (e.g., touch and sound) introduces challenges such as different sampling rates and data resolution. This thesis addresses the above challenges by contributing a human-centered framework for robot affordance learning that allows human teachers to guide the robot in the modeling process throughout the entire pipeline of affordance learning. We introduce several novel human-guided robot self-exploration algorithms that use human guidance to enable robots to efficiently explore the environment and learn affordance models for a diverse range of manipulation tasks. The work contributes a multisensory affordance model that integrates visual, haptic, and audio input, and a novel control framework that allows adaptive object manipulation using multisensory affordances.