Robot learning from demonstration using 3D computer vision

For robot application to proliferate in industry, and in unregulated environments, a simple means of programming is required. This paper describes methods for robot Learning from Demonstration (LfD). These methods used an RGB-D sensor for demonstration observation, and used finite state machines (FSMs) for policy derivation. Particularly, a method for object recognition was developed, which required only a single frame of data for training, and was able to perform real-time recognition. A planning method for object grasping was also developed. Experiments with a pick-and-place robot show that the developed methods resulted in object recognition accuracy greater than 99% in cluttered scenes, and manipulation accuracies of below 3mm in linear motion and 2° in rotation.