Dynamic collision estimator for collaborative robots: A dynamic Bayesian network with Markov model for highly reliable collision detection

In this paper, we propose a highly reliable and accurate collision estimator for robot manipulators working in human–robot collaborative environments, based on the Bayesian approach for practical uses. We assume the robot collision as a dynamic Markov process, not a static event, to reflect the transient behavior of mechanical collisions. Thus, the collision estimator can integrate the prior belief on collision and the measurements on the robot state, to produce the current belief on the robot collision in a usual recursive form. An exponential form of observation model, serving as the likelihood function, is constructed on the projected observation domain by using the multi-variate statistical information of empirical models of collision and non-collision cases. The proposed method is validated by using a commercial 7 degree-of-freedom (DOF) collaborative robot arm with random impacts along the links while it is in motion. Results show that the proposed collision estimator achieves a compelling performance with collision estimation time of 8.86ms on average, an overall accuracy of 99.47%, and zero occurrence of false alarm.