Sensing and perception for autonomous agents

点击翻译按钮获取中文摘要

Mobility is a key aspect of intelligence. Mobility enables an agent to explore its environment, to interact with its environment, and provides a realistic definition of here versus there. But mobility also introduces complex problems associated with sensing and perception. For example, mobility introduces fundamental questions such as; where is an agent in the environment? what is its pose? and questions related to change in pose: has an agent moved? and if so, how far and in what direction? What makes answering these problems difficult for an agent is that it rarely has access to some external process that can answer these questions directly. Rather the system (typically) must rely on sensors and sensing algorithms associated with the agent itself, and must integrate this information along with intent in order to obtain a best estimate for answers to these problems. Furthermore, getting answers to these questions wrong can have catastrophic consequences for the agent itself. For example, disorientation in divers, astronauts or pilots is a common problem leading to life-threatening consequences and mission failure. These fundamental problems are found across a range of disciplines, and not surprisingly have different names in different research cultures and when different sensors and sensor collections are used. For example, the task of estimating pose changes is found in egomotion estimation, motion from vection, structure from motion, and visual odometry. Similarly, robust solutions to the estimation of self-orientation and self-motion informs a wide variety of applications, from human performance in challenging environments, to virtual reality, to the design and development of autonomous vehicles. The primary research objective of this proposal is the development of robust and effective algorithms for motion and pose estimation for autonomous agents. Although biological and machine systems have access to a wide range of potential sensors and sensing modalities to address these questions, of particular interest here is in understanding how vision, and in particular binocular vision can be coupled with inertial sensing to estimate pose and motion. The basic approach to be followed in my research program is based upon a strategy that has proven particularly effective in other aspects of intelligent machines. Models of how biological systems solve similar tasks will be used to motivate solutions for machines. These solutions will then be evaluated in the lab and then in complex unstructured environments to test both the performance of the resulting system as well as to explore the efficacy of the underlying biological model.