<p>Deep reinforcement learning (DRL) has achieved state-of-the-art results in the field of autonomous driving within simulated environments, often surpassing human expertise. However, many existing DRL approaches rely heavily on track-specific features, limiting their ability to generalize across unseen environments. This study presents a track-agnostic reinforcement learning framework designed to enhance generalization in autonomous driving, by extending DRL to racing games without racetrack information. We evaluate two complementary DRL-based approaches: a simplified 2D racing environment utilizing sensor-based inputs (Lidar and accelerometer) and a highly realistic 3D racing simulation in TrackMania Nations Forever, incorporating both raw image pixels and physics-based data. We tested our trained models extensively on multiple tracks in each simulation. Our methodology leverages Proximal Policy Optimization to train agents capable of zero-shot transfer learning, enabling them to navigate previously unseen tracks without additional retraining. Experimental results indicate that our approach achieves competitive lap times while maintaining robust generalization capabilities. Furthermore, ablation studies and robustness tests demonstrate the framework’s adaptability to diverse driving conditions and stochastic perturbations. These findings contribute to the development of environment-independent reinforcement learning models for autonomous driving, with potential applications in real-world robotics and adaptive control systems.</p>