Tightly Coupled Sensor-Behavior Approaches for Real World Recognition

点击翻译按钮获取中文摘要

Tightly coupled sensor-behavior is important for intelligent systems to behave in the real world. Such systems should deal with uncertainty of sensor information, lack of information of the environment, uncertainty of motion, and uncertainty of a priori knowledge. Tightly coupled sensor-behavior is a promising approach to this problem. The result of this research is the followings :・ Interpretation of Sensor Information for Real World Understanding : We studied fusion of various visual information to recognize environments, and implemented the methods in J.Rob.Vision (developed in this research) to realize realtime processing. Considering the reliability of recognition results, we made the optimal plan generation method which minimize the cost of sensing and that of motions. Finally, we made experiments of understanding real environments by sensing and actions.・ Sensing mechanism for Behaviors Adaptive to the Real World : In order to move in a changing environment such as roads, we int … More egrated sonar-based obstacle detection, sign pattern-based recognition of roads and cars, and sign pattern learning during motion of a robot, and then we developed a guide robot for a blind people.・ Behavior Control Based on Environment Model and Sensory Information : By developing a three-finger manipulation system, we realized realtime detection of object posture by stereo vision and detection of grasping force on fingers. We proposed error avoidance by monitoring task states and verified by assembly experiments. Moreover, we realized a system which selects a suitable cloth in a pile of clothes by a recursive region split method, and picks it up by a hand with a special rotor.・ Adaptive Functions in Perception-Action Systems : We proposes "behavior network" which represents and controls concurrent process of selecting sensor-behavior and of predicting the environment based on the memory, and studied how to represent adaptive behaviors of a robot. We made a four legged robot with vision on J.Rob.Leg(developed in this research). By experiments of the adaptive behaviors in complex environments, we verified the performance of the behavior network. Less