Investigation of polyhedral shape representations and connectionist techniques in haptic object recognition

This paper presents two volumetric object representation models oriented to the dynamic integration of the information perceived by a robot hand in a haptic exploration. The enveloping polyhedral model is an upper approximation of the object derived from contact planes, whereas the approximating polyhedral model also takes advantage of the information provided by non-contacting hand elements. Features for object classification are derived from a spatial sampling of the polyhedral approximation. The effectiveness of these models in haptic recognition is investigated by considering alternative classification techniques and the haptic development of internal representations about the samples when formal models are unavailable. Experimental results demonstrate the effectiveness and robustness of the overall methodology