Path-following control of autonomous ground vehicles based on input convex neural networks

This paper studies the path-following problems in autonomous ground vehicles (AGVs) through predictive control and neural network modeling. Considering the model of AGVs is usually difficult to construct by first principles accurately, a data-driven approach based on deep neural networks is proposed to deal with the system identification tasks. Although deep neural networks have good representation capability for complex system, they are still hard to use for control area due to their nonconvexities and nonlinearities. Therefore, to make a trade-off between control tractability and model accuracy, the input convex neural networks (ICNNs) are developed to describe the dynamics of AGVs. As the designed neural networks are convex with regard to the inputs, the predictive control problem is converted to a convex optimization problem and thus it's easier to get feasible solutions. Besides, for adapting to different road conditions and some other disturbances, a periodically online learning algorithm is designed to update the neural network. Finally, two driving simulations under CarSim-Simulink platform are conducted to prove the superiority of our proposed techniques.

Automated summary

This paper investigates the path-following problems in autonomous ground vehicles (AGVs) using predictive control and neural network modeling. A data-driven approach based on deep neural networks is proposed to handle system identification tasks where the AGVs model is difficult to construct accurately. To balance control tractability and model accuracy, input convex neural networks (ICNNs) are developed to describe AGVs dynamics. A periodically online learning algorithm is also designed to adapt to different road conditions and disturbances. Two driving simulations are conducted to demonstrate the effectiveness of the proposed techniques.