End-to-End Learning vs. Modular Architectures: Comparative Insights into Autonomous Driving Systems
Autonomous driving system designers face a persistent dilemma: should they build a decomposed modular pipeline or train an end-to-end neural policy? This survey-style paper addresses that question by comparing modular, imitation-learning, reinforcement-learning, and hybrid architectures across interpretability, scalability, and robustness. Its central contribution is a practical twelve-criterion, four-dimensional framework intended to guide practitioners in selecting the right paradigm for a given deployment context.
This paper is a broad narrative review that culminates in a prescriptive architecture-selection framework. The comparative taxonomy is well-organized and accessible, but the work offers only lightweight empirical validation: the proposed framework achieves what the authors call seven correct or partial matches out of ten retrospectively chosen systems, using scoring conventions that are partly post-hoc. Because it openly abandons systematic-review protocols and contains no original experiments, its value lies in synthesis and practitioner guidance rather than in definitive or falsifiable conclusions.
The paper delivers a readable, historically grounded tour of autonomous-driving architectures from ALVINN to modern hybrid systems, accurately capturing the trade-offs between modular interpretability and end-to-end adaptability. The four-dimensional framework—safety and regulatory compliance, operating environment, data and computational resources, and deployment context—provides a sensible decomposition of real-world design constraints. Its discussion of sensing, perception, prediction, planning, and control modules in Section II is technically accurate and offers useful context for newcomers to the field.
The framework’s validation is methodologically thin. The ten-system benchmark mixes decades-old research prototypes with modern production stacks, and when the framework mispredicts ALVINN and PilotNet the authors simply rule them out-of-scope rather than refine the criteria.
The scoring logic is also inconsistently applied. Ambiguous cases such as TransFuser and UniAD are counted as partial successes by reclassifying end-to-end models with auxiliary losses as 'architecturally close to Hybrid,' conflating training objectives with system architecture. The criteria themselves are ad-hoc: C5 and C6 are described as signalling 'E2E' at the family level yet credited only to Hybrid in Table 4, and tie-breaking rules privilege Dimension 1 without theoretical justification.
All evidence is drawn from secondary literature via a non-systematic search across IEEE Xplore, arXiv, and Google Scholar. The authors candidly admit this limitation, noting the absence of PRISMA-style protocols and explicitly warning that quantitative metrics in their summary tables are incomparable across systems. Comparisons to prior surveys are descriptive rather than critical: the paper rarely highlights contradictions between existing reviews or adjudicates competing empirical claims. Consequently, readers seeking to understand which paradigm is actually safer or more scalable in controlled benchmarks will find limited new insight.
Because the paper contains no new experiments, hyperparameters and datasets are irrelevant; reproducibility instead hinges on whether an independent researcher can apply the decision framework and reproduce the 7/10 accuracy claim. This is unlikely. The criteria are stated in prose but no checklist tool, code, or adjudication rubric is released. Validation depends on the authors' subjective judgments—such as interpreting whether auxiliary-task end-to-end systems count as 'partial' Hybrid matches—and on tie-breaking conventions that are simply asserted. Without transparent classification records or inter-rater reliability, the framework currently functions more as an informed opinion structure than as a reproducible decision aid.
Autonomous driving systems have become a central focus of intelligent transportation research, with End-to-End Learning and Modular Architectures offering two prominent design paradigms for their implementation. E2E Learning uses deep learning algorithms to map raw sensory inputs directly to driving actuators, providing a streamlined and adaptable solution. while Modular Architectures employ a pipeline-based approach, dividing the system into distinct subsystems for perception, cognition, planning, and control. This paper presents a comprehensive comparative analysis of these paradigms, focusing on their strengths, limitations, and trade-offs to provide insights into their suitability for various autonomous driving applications. The study evaluates key factors such as interpretability, scalability, robustness, and real-world applicability. While End-to-End Learning emphasizes simplicity and adaptability in dynamic environments, it lacks transparency and is highly dependent on large datasets. Conversely, Modular Architectures offer superior interpretability and task-specific optimization, but face challenges related to integration complexity and scalability. To address these limitations, hybrid approaches that combine the strengths of both paradigms have emerged, offering a promising direction for overcoming these challenges. Beyond this comparative synthesis, following work proposes a Four-Dimensional Architecture Selection Framework, comprising twelve binary criteria across safety, operating environment, data/computational resources, and deployment context, and validate it against ten published autonomous driving systems, correctly recommending 7/10 deployed architectures. This work synthesizes existing literature to highlight key trade-offs between the paradigms and identifies hybrid architectures as a promising direction for future research.
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.