OA-PINN: Efficient Obstacle Avoidance for Autonomous Vehicle Safety with Physics-Informed Neural Networks

The safe operation of Autonomous Vehicles (AV) primarily relies on effective collision avoidance techniques. Therefore it is essential to integrate obstacle avoidance features into the motion planning of vehicles. This research focuses on obstacle avoidance using the CBF which has evolved as an efficient mathematical tool for unmanned vehicles to ensure safe navigation. A Quadratic Programming (QP) problem for collision avoidance is formulated using CBF as a constraint. The Hamilton-Jacobi-Bellman (HJB) equation derived for the QP problem gives rise to a Partial Differential Equation (PDE), the solution of which generates the necessary control input for the vehicles to successfully avoid the obstacles. The key novelty of this research lies in solving the HJB equation using a multilayer perceptron, called a Physics-informed Neural Network (PINN), which needs less computation in a cluttered environment. The proposed methodology integrates the safety and reliability aspects of CBF with an obstacle avoidance solution using PINN (OA-PINN). The performance of OA-PINN is validated using numerical simulations and related hardware experiments.