Autonomous Learning for Interactive Agents

Robots that interact with humans are still confined to controlled spaces, such as lab environments, where they conduct highly pre-specified task in interaction with recruited and cooperate users. Some of the obstacles that restrict real world applicability (amongst others) are their heavy reliance on domain-specific pre-programming and learning tasks that arise from the real world rather than being contrived for the purpose of robot training. In this talk, I will present a research direction on autonomous learning that aims to alleviate the above problems and push interactive robots over the edge of wider usability. The core of my research lies in multi-task reinforcement learning that helps agents to understand and optimise their behaviour by interacting with humans and learning from feedback and examples. I will present three applications of this autonomous learning framework: (1) a situated agent that learns to guide people in complex indoor environments using a divide-and-conquer approach, (2) a conversational robot that learns to play educational games from interacting with children, and (3) a strategic agent that learns trading negotiations using deep reinforcement learning. I will conclude by discussing directions for future research that further increase the level of autonomy of interactive agents in the real world.