Model-Based Planar Contour Following in the Presence of Pose and Model Errors

Abstract

This article shows how the use of a geometric model during planar contour following with a force-controlled robot results in a higher quality (i.e., faster and/or more accurate) task execution, even in the presence of pose and model errors. Neither the pose (i.e., position and orientation) of the contour nor the starting point on the contour have to be known in advance. The contour is described by curvature and/or orientation of the normal direction, as a function of arc length. The con tour is identified on-line from the sensor measurements, and matched on-line with its model. This yields an estimate of the point in the model that corresponds to the real contact point. The curvature from the model is then used to generate feed- forward velocity set points. The article compares a nonrecursive deterministic least- squares matching strategy with recursive stochastic matching strategies, based on the extended Kalman filter and the Page- Hinkley hypothesis test. Matching is preferably based on cur vature profiles, since these are Euclidean invariant and hence reduce the search space of the matching. On contours with piecewise constant curvature, however, the extended Kalman- filter strategy uses the orientation of the normal direction. Extensive experimental evaluation of these strategies is given, including shape matching based on data during a deburring task. 1. The definition of this tangent frame is based on the Frenet frame; see Section 2.1. 2. Curvature, arc length, and orientation of the normal are defined in Section 2.1.