Detection and localization of unmodeled manipulator collisions

Abstract

Robotic tasks usually require some collision free motions, and there has been considerable work in methods for collision avoidance. However, noise in the sensor data, movement of the obstacles, and incomplete or inaccurate model of the surroundings all may lead to unexpected collisions. Detecting such collisions is necessary before recovery and/or replanning may take place. A means of detecting a collision, as well as the position of the collision on the manipulator has been developed. The detection scheme combines information from observed disturbance torques to detect collision and infer the location of contact with the environment. Knowledge of contact position allows for a more intelligent and less error-prone recovery scheme. A simulation using a three DOF manipulator shows that the collision identification and localization scheme is feasible and robust with respect to noise.