Robust Reinforcement Learning for Autonomous Driving in Uncertain Environments

Autonomous vehicles typically operate in highly dynamic and uncertain environments where conventional reinforcement learning methods often fail to adapt due to fixed transition dynamics, static reward structures, and Markovian state assumptions. These limitations lead to environment-specific overfitting, reducing policy robustness in unseen conditions. To address this challenge, we propose a history-dependent reinforcement learning framework that integrates randomized training dynamics, context-aware reward functions, and memory-based policy optimization using Long Short-Term Memory (LSTM). Our method enables agents to implicitly infer latent environmental variations, dynamically adjust decision-making priorities, and improve robustness across diverse driving conditions without requiring explicit access to environment parameters. We employ Proximal Policy Optimization (PPO) with an LSTM-based policy architecture, leveraging historical interactions to encode environment-aware representations and enhance policy adaptability. The proposed method requires no fine-tuning and supports real-time decision-making, making it practical for deployment in real-world driving applications. Experiments performed in CARLA across various scenarios demonstrate the effectiveness of our approach in achieving reliable decision-making in adaptive cruise control systems.