Learning from Human Driver Demonstration for Speed Control of Ground Autonomous Vehicles

Recent developments in autonomous driving aim to emulate human decision-making, offering potential benefits for safety and user acceptance. Nevertheless, a significant challenge remains in accurately modeling the complex and adaptable nature of how humans navigate the roads. This work introduces a control system that learns from human driving behavior. It utilizes FIFO buffering to process a series of inputs and LSTM networks for making predictions. The inputs include historical buffered data, current sensor readings, and predicted averages of chunked pitches aimed at estimating axle torque and deceleration values. Validation tests using data from sensors installed on a Chevy Bolt EUV have confirmed the predictive accuracy of the system, favorably compared with ground truth measurements. Moreover, its ability to filter out additive noise with zero mean from both current and predicted averages of chunked pitch values makes it less susceptible to additive noises. Statistical analyses, including measures of central tendency, coefficient of correlation, mean absolute error and probability density plots have further validated the approach. These analyses demonstrate the ability of the proposed method to replicate human-like decision-making in driving, underscoring its potential to enhance autonomous driving. By ensuring safety, comfort, and efficiency, the proposed method closely mirrors the responsiveness of a human driver.