Interactive autonomous machines

点击翻译按钮获取中文摘要

Although there have been significant strides in our ability to design and deploy autonomous systems, key problems remain unsolved that are critical to future advances in this field. My research program concentrates on three of these key problems: multi-cue integration, plan development and reasoning, and human-robot interaction. Multi-cue integration: An autonomous machine must be able to capture sensory data from a range of different technologies to build a model of where it is and how it can move within its environment. The underlying approach here is to understand how humans build similar representations and to leverage this knowledge to solve the problem for machines. Over the last ten years or so I have been working with a group of international collaborators to understand how humans develop a sense of self-orientation and self-motion. Key here is performing controlled experiments in environments within which humans are presented with unusual cue combinations to illuminate the underlying cue integration process. With the completion of recent work in long duration bed rest, I have helped develop a deep and wide dataset of human self-orientation perception which I plan to use to construct a time-dependent cue integration model that will be evaluated using underwater robots. Plan development and execution: The machine must be able to reason about its environment and how it can act upon that environment. Traditionally, long term robot plans were developed using mechanisms that decomposed tasks based on logical structures. Although effective in its day, these planning approaches have been eclipsed by modern AI approaches (e.g, Deep Reinforcement Learning - DRL) that have demonstrated extraordinary capabilities, especially for tasks with reasonably short-term horizons for which appropriate training can be performed. How can we best integrate the lessons learned from classic planning regimes with the capabilities of mechanisms like DRL? Utilizing a set of specific tasks - invasive water plant monitoring and indoor environment monitoring - I plan to explore how best to use DRL to learn short-term plans that can be sequenced to provide task solutions for these complex, but highly structured tasks. Human-robot interaction (HRI): The machine must be able to interact with the environment and most critically it must be able to interact with humans that occupy the environment within which it is operating. My preliminary work suggests task-specific gesture language can be effective for HRI, but can we better leverage the underlying task structure to build effective HRI systems? Utilizing common commercial diver tasks, I plan to generalize my preliminary work here to condition language token and conversational structure priors within the HRI representation, and to leverage both in understanding human to robot communication.