A Survey on End-to-End Autonomous Driving Training from the Perspectives of Data, Strategy, and Platform
End-to-end autonomous driving maps raw sensors directly to control actions, yet its real-world bottleneck lies in the training ecosystem rather than model architecture alone. This survey introduces a Data-Strategy-Platform taxonomy that frames training as an interdependent system: data defines what is learnable, strategies govern how learning aligns with driving objectives, and platforms enable scalability. The authors argue the field must shift from raw data volume to value density, from isolated task-specific learners to universal foundation models, and from static offline training to closed-loop training-testing integration.
The paper delivers a competent, wide-ranging survey organized around a three-layer taxonomy. Its primary value is structural: it brings data pipelines, learning paradigms, and compute infrastructure into one coherent picture. However, the analysis rarely advances beyond description; the interactions among layers are asserted rather than formally modeled, and the forward-looking vision reads as a sensible concatenation of existing desiderata rather than a sharply defined agenda. As a reference map it is useful, but readers seeking novel theoretical synthesis will find the framing thinner than the breadth suggests.
The survey’s scope is genuinely comprehensive, covering data valuation, imitation learning, RL, diffusion policies, MLLMs, world models, distributed training, and industrial deployment practices. Particularly useful are its comparative distillations: Table III contrasts five training paradigms by multimodality, autoregressive error, trajectory-level generation, and interpretability, while the discussion assigns each paradigm a specific 'best role' instead of treating any as a universal solution. The authors also candidly acknowledge the speculative nature of industrial practice descriptions.
The Data-Strategy-Platform taxonomy is coherent but thinly formalized: inter-layer coupling is described in broad conceptual strokes rather than with a testable model. Equation (1), $V(s)=w_{r}\cdot R(s)+w_{u}\cdot U(s|\theta)+w_{d}\cdot D(s|\mathcal{D}_{train})$, proposes a linear composite for scenario value yet offers no procedure to learn or calibrate its weights, and no empirical validation on any dataset. The industrial comparison in Table V is openly based on 'publicly available descriptions,' which the authors admit 'should be read as indicative patterns rather than exhaustive descriptions of proprietary systems,' substantially weakening its evidentiary authority. Additionally, the distributed training discussion in Section V-A is generic large-model infrastructure advice with little tailoring to the safety-critical, latency-bound multimodal constraints specific to autonomous driving.
The evidence consists of a broad citation map spanning datasets, architectures, simulators, and systems papers. Table I fairly positions this survey against prior reviews by highlighting its unique coverage of data, strategy, and platform together. Yet the methodology is opaque: there are no reproducible inclusion criteria, literature search protocols, or coding boundaries, so coverage may be opportunistic rather than systematic. While the paradigm-level trade-offs in Table III are well structured, the industrial claims in Table V carry weak verifiable authority because they are explicitly acknowledged as reconstructed from public relations sources and second-hand reports.
As a conceptual survey, reproducibility rests on bibliographic transparency and the authors' commitment to a living repository. The paper notes that it 'maintain[s] a continuously updated repository tracking cutting-edge literature,' which supports community revisiting of the source material. However, there is no systematic review protocol—no query strings, PRISMA-style selection criteria, or annotation methodology—making the survey process itself difficult to reproduce. The value-scoring framework in Equation (1) is not instantiated or benchmarked, and the industrial practices table cites no verifiable technical source for specific platform claims, leaving independent verification largely impossible.
Autonomous driving is a cornerstone technology for the future of intelligent transportation, where end-to-end learning has emerged as a transformative paradigm that directly maps multimodal sensory inputs to driving actions through unified differentiable models. While offering advantages, the effectiveness of end-to-end autonomous driving (E2E-AD) is ultimately determined by the quality of its training ecosystem. This paper provides a comprehensive review of training methods and ecosystem for E2E-AD. We introduce a Data-Strategy-Platform taxonomy that conceptualizes training as an interdependent system. The data layer defines what can be learned, the strategy layer governs how learning aligns with driving objectives, and the platform layer supports scalability and continuous evolution. Within this framework, we survey recent advances across data-centric pipelines, learning paradigms, and training infrastructures, and analyze their interplay in shaping model performance, robustness, and deployability. Finally, we reflect on current limitations and articulate a forward-looking vision that emphasizes a shift from data quantity to data value, from isolated optimization to foundation-driven generalization, and from static training to integrated training-testing loops, aiming toward robust, scalable, and trustworthy autonomous driving systems. We maintain a continuously updated repository tracking cutting-edge literature and works at \href{https://github.com/Jiaaqiliu/Awesome-Training-Ecosystem-for-E2E-AD}{Our Project Page}.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.