Reinforcement Learning-Based Post-Training Optimization for End-to-End Autonomous Driving Models

Abstract

End-to-end autonomous driving models increasingly map raw sensor observations directly to control actions, promising reduced engineering complexity and improved adaptability relative to modular pipelines. However, these models remain vulnerable to distribution shift, rare safety-critical events, and reward misalignment when deployed in open-world traffic. This paper examines reinforcement learning-based post-training as a system-level optimization layer for end-to-end driving models. It argues that post-training is not merely a model refinement technique but a socio-technical intervention that reshapes architecture, data infrastructure, validation regimes, governance, and public accountability. The discussion situates imitation-based pre-training alongside reinforcement learning fine-tuning, preference modeling, and constrained policy optimization. It analyzes structural trade-offs involving simulator fidelity, compute intensity, latency constraints, safety envelopes, and regulatory oversight. The paper further considers fairness across geographic contexts, sustainability implications of large-scale training, and policy mechanisms for staged deployment and incident reporting. A central claim is that effective post-training requires an integrated stack in which reward design, human feedback, formal safety constraints, and deployment monitoring are co-engineered. Without such integration, reinforcement learning post-training may improve aggregate performance while amplifying opaque failure modes. The paper concludes with a research agenda for auditable, resource-aware, and socially legitimate post-training pipelines for autonomous driving.