Acoustical Recognition of Objects in Robotics II. Determination of Type, Pose, Position, and Orientation

Abstract

Optical recognition of objects in a robotics context encounters many sources of difficulty (e.g., polished objects, dark objects, extraneous light sources, etc.). However, acoustical recognition, i.e., recognition based upon the interpretation of scattered acoustical waves, has some intrinsic advantages, the most important of which is that in the absence of noise, the scattered waves depend only upon the external geometry and acoustical impedance of the object and the table that supports it. Specifically, our problem is to deduce the state of the object (i.e., type, pose, horizontal position, and azimuthal orientation) given a set of acoustical, pulse-echo, scattering data representing a sufficient diversity of directions and temporal frequencies. We use a decision-theoretic approach to the recognition problem, i.e., we determine the most probable state given the measured scattering data. A central element is a measurement model embracing certain statistical submodels and a representation of the scattering process. Here, we use an experimental approach to providing such a representation, thereby avoiding the errors in simple scattering theories (e.g., Kirchhoff). To illustrate the effectiveness of this approach, we have tested the recognition algorithm on real experimental scattering data obtained from a set of 4 objects with unknown states. The results were very satisfactory and constitute a proof of principle.

This work was supported by the Independent Research and Development Funds of Rockwell International.