A real-time human-perception interface for task-level control of a robot in unfamiliar environments.

摘要

Recent advances in the development of semi-autonomous robotic systems offer numerous potential advantages in many engineering and science endeavors. Significant reductions in cost, time and risk, as well as increased capability, can be obtained by utilizing intelligent machines to assist humans. However, the use of robots also introduces many challenging issues, including the need for high-bandwidth stable control despite communication delays and operator fatigue. In response to these challenges, the Stanford Aerospace Robotics Laboratory has pioneered the Task-Level Control architecture, which enables humans to direct, from a strategic level, sophisticated tasks that a robot then executes autonomously.;The research reported here is intended to extend the Task-Level Control architecture significantly--by using human perception in a natural way--to work well in unfamiliar environments. An unfamiliar environment is defined to be one about which it is impossible to have perfect and complete knowledge before developing and deploying a robotic system. Clearly, every work environment is, to some extent, unfamiliar. This research has shown that drawing intimately, in real time, upon a human's deep visual perception is extremely effective in overcoming such unfamiliarity.;A novel interactive vision-based operator interface for directing a highly autonomous robot operating in an unfamiliar environment is presented. Intuitive interaction with a live-video display from cameras on board the robot is used in combination with stereo-vision algorithms to maintain the operator's attention at the overall object-level during the modeling process. With this interface, the human's remarkable ability to discern entire object-level constructs is utilized to produce quick, cogent and robust models of unexpected and unknown objects in the environment.;Once unfamiliar objects have been suitably modeled, tasks involving those objects can be directed via the Task-Level Control architecture. Utilizing on-board sensing, low-level dynamic-control autonomy, strategic logic, and path-planning algorithms, the Task-Level Control architecture enables an operator to request effortlessly sophisticated, object-based tasks which the robot then executes autonomously. Tasks such as robot navigation and the capture and manipulation of previously unfamiliar, moving objects in a newly-modeled, obstacle-cluttered environment have been successfully demonstrated. Experimental results with a free-flying laboratory robot are presented.