A Survey on Imitation Learning Techniques for End-to-End Autonomous Vehicles

Le Mero, Luc; Yi, Dewei; Dianati, Mehrdad; Mouzakitis, Alexandros · 2022 · Crossref

DOI: 10.1109/tits.2022.3144867

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper presents a comprehensive survey of Imitation Learning (IL) techniques applied to end-to-end autonomous vehicle systems. Motivated by the industry’s shift from manually designed modular architectures—which suffer from computational inefficiency and error propagation—to self-optimizing end-to-end systems, the authors address the challenge of training vehicles to map raw sensory inputs directly to control signals. While end-to-end systems offer computational efficiency, they face significant hurdles regarding verifiability, generalizability, and the "cascading error" problem, where small prediction errors compound over time. The study aims to categorize the state-of-the-art literature, evaluate available datasets and simulators, and identify open challenges to facilitate future development in this safety-critical domain. The authors classify Imitation Learning into three distinct sub-fields based on algorithmic approaches: Behavioural Cloning (BC), Direct Policy Learning (DPL), and Inverse Reinforcement Learning (IRL). BC is treated as a supervised learning task that maps states to actions, though it struggles with distributional shifts. DPL utilizes iterative online training by querying an expert for optimal solutions in states encountered by the agent. IRL attempts to infer the underlying reward function from expert demonstrations to train an agent. The review focuses heavily on BC, further subdividing it into end-to-end control prediction, direct perception, and uncertainty quantification. The authors also provide a comparative evaluation of datasets and simulation tools, such as CARLA, which are essential for training these data-intensive deep learning models. Key findings highlight the evolution and limitations of current IL methods. Early systems like ALVINN and DAVE demonstrated the viability of end-to-end lane following and obstacle avoidance. More recent works, such as those by Bojarski et al., utilize multi-camera inputs to improve steering control, while studies by Cultrera et al. and Hecker et al. introduce attention mechanisms and 360-degree sensor inputs to enhance model explainability and performance. The survey notes that Conditional Imitation Learning (CIL), which incorporates high-level intent commands, significantly improves robustness against perturbations. However, the authors identify critical limitations: models trained via BC often fail to generalize to scenarios not present in the training data, and pure end-to-end control prediction struggles with complex navigation tasks requiring turns or dynamic decision-making. Furthermore, the lack of interpretability in deep learning models remains a barrier to safety verification. The significance of this work lies in its structured categorization of IL techniques, providing a clear framework for researchers to understand the trade-offs between different approaches. By reviewing the history, problem formulations, and specific implementations of BC, DPL, and IRL, the paper underscores the necessity of addressing generalizability and safety. The authors conclude that while IL offers a promising path toward fully autonomous vehicles, future research must focus on improving model robustness, integrating uncertainty quantification, and developing methods that can reliably handle the complex, dynamic nature of real-world driving environments. This survey serves as a foundational reference for advancing end-to-end autonomous driving systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• driverless ads