GoonDAE: Denoising-Based Driver Assistance for Off-Road Teleoperation

Abstract

Due to the limitations of autonomous driving technology, teleoperation is used extensively in hazardous environments such as military operations. However, the performance of teleoperated driving is primarily influenced by the driver's skill level. In other words, unskilled drivers need extensive training for teleoperation in harsh and unusual environments, such as off-road. In this letter, we propose GoonDAE, a novel denoising-based real-time driver assistance method that enables stable teleoperated off-road driving. We introduce a denoising autoencoder (DAE) based on a skip-connected long short-term memory (LSTM) to assist the unskilled driver control input through denoising. In this approach, it is assumed that the control input of an unskilled driver is equivalent to that of a skilled driver with noise. We train GoonDAE using the skilled driver control inputs and sensor data collected from our simulated off-road driving environment. Our experiments in the simulated off-road environment show that GoonDAE significantly improves the driving stability of unskilled drivers.