Towards High Accuracy Robot-Assisted Surgery

Abstract

In this article, we propose a new error modeling approach for robot manipulators in order to improve the absolute accuracy of a tools pose by using polynomial regression method. The core idea is based on a well-known fact: accuracy of repeatedly reaching a given position is much higher than the accuracy of absolute positioning (i.e. moving the manipulator to a given position). The underlying reason is, that positioning errors are dominated by systematic errors - while stochastic errors are significantly smaller. This fact is then exploited to apply some learning algorithm to derive an error-compensation model. Technically, this means that the robot is being calibrated a priori with some external sensor once. Afterwards it can operate with a much better quality. In detail, we propose to first perform a coordinate transformation using a least mean square approach (for registration). Then, to account for deviations of measured position in comparison to nominal (robot) position, the frame transformation model at each robots joint is extended by translational and rotational error parameters. This is then used to built an error compensation model with regression techniques. We evaluate the method on a data set obtained using a 7DOF robot manipulator and show that this approach brings positioning error to the order of repeatability errors for this manipulator.