Autonomous Vehicle Localization with Deep Reinforcement Learning

• Henrique Barros Oliveira USP
• Felipe Gomes de Melo D'Elia USP
• Larissa Driemeier USP

Resumo

Accurately estimating a vehicle's position is a key challenge in autonomous navigation, requiring real-time performance and robustness to sensor noise, outages, and unmodeled dynamics. Recent advances in deep reinforcement learning (DRL) have shown promise as a data-driven, model-free alternative for various robotics and autonomous driving tasks. This work proposes an end-to-end DRL-based pipeline for vehicle localization using only proprioceptive sensor data, including IMU, speed, and steering angle, combined with estimates from an Extended Kalman Filter. The proposed system is built around a containerized architecture that enables simplified deployment from training in the CARLA simulator to real-time execution on embedded hardware. The approach is validated both in simulation and on a 1: 10 scale physical vehicle, demonstrating stable training convergence and real-time inference capabilities on resource-constrained platforms. While the model's accuracy is not yet sufficient for closed-loop operation, the results confirm the viability of the architecture and the training approach as a foundation for further development.

Palavras-chave: Training, Location awareness, Accuracy, Computational modeling, Pipelines, Computer architecture, Deep reinforcement learning, Real-time systems, Vehicle dynamics, Autonomous vehicles, autonomous vehicles, self-localization, reinforcement learning