Experimental Study of Tracked Vehicle Velocity Using Estimated Disturbance and Machine Learning for Application to Environments Different from Those in Training

Abstract

This study examines the training policies and environmental robustness of a neural network used in velocity estimation for a tracked vehicle with slippage. In the proposed method, the velocity is estimated by a neural network whose input is an estimated disturbance to the driving axle that includes slippage information. First, we experimentally clarify the proposed method's scope of applicability and effectiveness under different environmental conditions in training and estimation. Subsequently, we experimentally confirm that the estimated disturbance is robust to environmental changes and complementary to environmental information. Finally, the neural network trained on a flat surface is validated in combination with gravity compensation for acceleration to apply it to driving on a slope.