Object classification for robot manipulation tasks based on learning of ultrasonic echoes

We describe an object recognition technique based upon the extraction of simple features from the initial part of ultrasonic echoes. Features collected from a single or multiple viewpoints are classified using a decision tree. Since only the initial part of the echo is examined, the approach has potential for faster classification than alternative techniques requiring processing of the entire waveform. To emulate a workcell scenario, the approach has been verified mounting a Polaroid sensor at the wrist of a Puma 560 manipulator and implementing a simple modification of the proprietary circuitry (Polaroid Ranging Unit 6500). When tested with a set of 8 small plastic objects with regular shapes, the recognition technique has achieved classification success rates from 72% to 98%, depending upon the number and selection of echoes exploited for recognition. The paper illustrates classification performance using single or multiple viewpoints under both axis parallel and oblique decision trees.