CAREER: Robot Reflection in Lifelong Adaptation

点击翻译按钮获取中文摘要

Lifelong adaptation will enable robots to operate over long periods of time in unstructured and ever-evolving worlds. Without effective adaptation, lifelong robots will inevitably fail, jeopardizing both their missions and surrounding environments. Change and novelty are two of the greatest challenges that prevent robots from achieving lifelong adaptation. The goal of this award is to develop and gain a fundamental understanding of a novel paradigm, called robot reflection, to address change and novelty in order to achieve lifelong robot adaptation. Inspired by theories of cognitive reflection for human adaptation, the proposed robot reflection paradigm will allow a robot to reason about itself and adjust its own adaptation process without requiring human supervision. This award stands to provide significant benefit to society by enabling robots that reflect to improve performance over their lifetimes, thus broadly impacting lifelong robotics applications in which robots must operate over long periods of time, such as search and rescue and autonomous driving.To produce algorithms for robot reflection in lifelong adaptation, the technical aims of this award are divided into three thrusts. The first thrust enables self-aware robot adaptation in context, through a principled method that offsets a robot perception and decision making model according to combinations of changes in the environment and the robot itself. The second thrust addresses adapting robot adaptation according to the degree of novelty in the robot's experiences, through a novelty estimation approach that compares the current experiences against past episodes in the robot's memory to estimate novelty, and a machine forgetting method to update the robot's fixed-capacity memory by adaptively selecting past episodes to forget while avoiding catastrophic forgetting. The third thrust focuses on robot adaptation by imagination without requiring a robot to physically experience a new scenario, through a novel approach that formulates robot imagination as sequential decision making and explores interpretable methods for modeling the underlying relationships between experienced and new scenarios for data-efficient imagination.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.