Self-Supervised Post-Training for Robust End-to-End Vehicle Control

Abstract

End-to-end vehicle control has emerged as a compelling alternative to modular autonomous driving stacks because it can learn perception, prediction, planning, and actuation within a unified differentiable architecture. Yet the same integration that enables graceful behavior in nominal conditions also creates fragility under distribution shift, sensor degradation, rare edge cases, and evolving road environments. This paper examines self-supervised post-training as a system-level strategy for improving the robustness of end-to-end vehicle control after initial supervised or reinforcement-learning-based training. Rather than treating self-supervision as a mere pretext for representation learning, the discussion positions post-training as a socio-technical process that couples model adaptation, fleet data infrastructure, verification, governance, and lifecycle sustainability. The analysis considers architectural trade-offs between monolithic policies and modular safety envelopes, the role of large-scale unlabeled driving data, and the operational challenges of continuous over-the-air updates. It also explores fairness, privacy, and policy implications when self-supervised adaptation depends on fleet telemetry and simulation pipelines. The central argument is that robust end-to-end control cannot be achieved through algorithmic innovation alone; it requires disciplined post-training regimes, auditable data provenance, safety-case integration, and deployment governance that treats model updates as safety-critical changes. By connecting representation learning with infrastructure and regulation, the paper outlines an interdisciplinary agenda for self-supervised post-training that is technically effective, operationally accountable, and socially sustainable.