On the robustness of haptic object recognition based on polyhedral shape representations

Polyhedral models derived from haptically-sensed information have been shown to effectively support recognition of convex objects in known poses. In this paper we investigate the robustness of the recognition methodology based on these models with respect to various sources of potentially inaccurate knowledge, including inexact a priori knowledge about shape and location of the objects to be identified and noise in the sensed information. The methodology is shown to exhibit some general robustness property, essentially stemming from the underlying volumetric modeling approach. Specific attention is paid to the effect of errors in the sensed normal directions at the contacts, since reliable estimation of this quantity could prove difficult in the practical application of the methodology. To this purpose, a technique is developed for identification and pruning of outliers from the data set. Further, a polyhedral model is defined which does not rely on normals but only on contact localization and hand pose information. Experimental results assess the impact on recognition performance of the loss of the normal information