GRAND: Graph Laplacian-based Reinforcement Learning for Autonomous Navigation Decisions

Autonomous Navigation helps make navigation efficient by taking smart decisions while prioritising both safety and smoothness. The primary challenge encountered is developing a system that makes real-time decisions accurately. This system should be able to recognise and classify several different objects encountered and take appropriate action based on the situation. The act of driving can be viewed as a sequence of decisions taken with occasional feedback from the environment. This makes reinforcement learning an appropriate algorithm for autonomous navigation. The proposed approach makes use of periodic sampling to collect data from a simulated environment and then computes the feature matrix using the graph Laplacian approach. This feature matrix is then used to fix policy weights for each objective in an iterated policy approximation fashion. Greedy optimal policy computation and weighted voting are then used to compute the final action to be taken. This paper proposes GRAND, a system that tackles the challenges involved in making a decision when the ego vehicle encounters unexpected turns in a simulated environment with multiple paths. The results illustrate the advantages of using the graph laplacian approach along with the MO-API algorithm in different road layouts. Furthermore, this approach can be applied to real vehicles for increased safety and efficiency. This would reduce traffic congestion and provide a more reliable method for public transportation, delivery systems, and even military applications.