Vision Processing for Robot Learning

Robot learning | be it unsupervised, supervised or self-supervised | is one method of dealing with noisy, inconsistent, or contradictory data that has proven useful in mobile robotics. In all but the simplest cases of robot learning, raw sensor data cannot be used directly as input to the learning process. Instead, some \meaningful" preprocessing has to be applied to the raw data, before the learning controller can use the sensory perceptions as input. In this paper, two instances of supervised and unsupervised robot learning experiments using vision input are presented. The vision sensor signal preprocessing necessary to achieve successful learning is also discussed. 1 Background By virtue of their ability to change location, mobile robots are exposed to noisy, inconsistent, or contradictory sensory perceptions, to a larger degree than, for instance, stationary robots. Any sensor signal processing mechanism that can cope with such data is therefore particularly suitable for mobile robot control. Artiicial neural networks have been shown to be a suitable mechanism for this purpose (e.g. Nehmzow 92, Zalzala & Morris 96]). In this paper, we discuss two experimental scenarios, in which a Nomad 200 mobile robot (see gure 1) acquires fundamental sensory-motor competences through neural network learning, using input from a CCD camera. In the rst example, supervised teaching is used to train the robot to avoid obstacles, follow walls and traverse corridors. Robot training is a fast and reliable control method ((Nehmzow 95b]), and the experiment presented here demonstrates that it can successfully be applied to vision data. In the second experiment, vision data is clustered autonomously by the robot, using unsu-pervised neural network learning, to diierentiate between images containing boxes and those not containing boxes. Both experiments share similar image preprocessing procedures, and would most probably not have succeeded without those preprocessing methods. The paper discusses these vision data preprocessing methods in detail. All experiments reported here were conducted at Manchester University, using a Nomad 200 mobile robot called FortyTwo (see gure 1). The robot is a hexagonal robot, 50 cm in diameter and 80 cm high (weight 59 kg). It is equipped with sixteen sonar range nding sensors (range 15 cm to 6.50 m), sixteen infrared