Online learning terrain classification for adaptive velocity control

Abstract

Safe teleoperation during critical missions, such as urban search and rescue, and bomb disposal, requires careful velocity control when different types of terrain are found in the scenario. This can particularly be challenging when mission time is limited and the operator's field of view affected. This paper presents a method for online adapting robot velocities according to the terrain classification results combined from vision- and laser-based classifiers. The vision-based classifier is self-supervised and adapts itself according to the vibration sensing and the pose estimation of the robot. The image patches where the vibration data are gathered are used to train the vision-based classifier. The Support Vector Machine is used for the laser-based classifier to train and classify the data. The final prediction result is produced by using the Naive Bayes Classifier to fuse the vision- and laser-based classifiers. The system is robust to illumination variations, and can be improved online given feedback from the operator.