Human-Aided Online Terrain Classification for Bipedal Robots Using Augmented Reality

This paper presents an online training system, enhanced with augmented reality, for improving real-time terrain classification by humanoid robots. The real-time terrain type prediction model relies on data acquired from four different sensors (force, position, current, and inertial) of the NAO humanoid robot. We compare the performance of Stochastic Gradient Descent, Passive Aggressive classifier, and Support Vector Machine in predicting the terrain type being traversed. Then, the models are trained online by manually inputting the correct terrain type being traversed to improve the accuracy of the predictions over time. An Augmented Reality (AR) user interface is designed to display the robot diagnostics and terrain type being predicted and obtain the user feedback to correct the terrain type when needed. This allows the user to improve the classification results and enhance the data collection process in the easiest way possible. The experimental results show that the Passive Aggressive classifier is the most successful among the three online classifiers with an accuracy of 81.4%.