Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
102 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

A Survey of End-to-End Driving: Architectures and Training Methods (2003.06404v2)

Published 13 Mar 2020 in cs.AI and cs.RO

Abstract: Autonomous driving is of great interest to industry and academia alike. The use of machine learning approaches for autonomous driving has long been studied, but mostly in the context of perception. In this paper we take a deeper look on the so called end-to-end approaches for autonomous driving, where the entire driving pipeline is replaced with a single neural network. We review the learning methods, input and output modalities, network architectures and evaluation schemes in end-to-end driving literature. Interpretability and safety are discussed separately, as they remain challenging for this approach. Beyond providing a comprehensive overview of existing methods, we conclude the review with an architecture that combines the most promising elements of the end-to-end autonomous driving systems.

End-to-End Driving: Architectures and Training Methods

This paper presents an exhaustive survey of end-to-end approaches to autonomous driving, offering a detailed analysis of mechanisms where the complete driving pipeline is represented by a single neural network. While traditional modular approaches partition the pipeline into perception, localization, planning, and control, the end-to-end paradigm inherently simplifies the system architecture. However, this simplification introduces intricate challenges related to interpretability and safety, both of which are addressed in the discourse.

Comparative Analysis of Driving Approaches

The paper delineates the distinction between modular and end-to-end approaches. The modular approach, characterized by interconnected yet autonomous modules, benefits from interpretability—a fault in the system can be traced to a specific module. Conversely, end-to-end systems, by optimizing the entirety of the driving task in a unified learning process, promise a reduction in complexity, although at the cost of diminished interpretability. The complexities of environment representation and decision-making required in modular approaches are bypassed in end-to-end models, albeit posing challenges in situations demanding fine-grained fault analysis.

Architectures and Learning Paradigms

Diverse architectures are surveyed, including convolutional neural networks (CNNs) for perception, and recurrent neural networks (RNNs) for temporal sequence analysis. The paper extensively reviews learning frameworks suitable for end-to-end driving, emphasizing imitation learning (IL) and reinforcement learning (RL). IL, with its rooted technique of deriving driving actions from expert drivers, contends with the distribution shift problem wherein driving errors lead to trajectories that diverge from those experienced during training. The paper proposes solutions such as data augmentation and on-policy learning to mitigate these issues. In contrast, RL is noted for its potential to explore novel states, although it often requires extensive simulation due to its data inefficiency.

Input and Output Modalities

End-to-end models leverage various sensor inputs, including monocular and stereo vision, LiDAR, and high-definition maps, often combining them to enhance robustness. The integration of navigational commands enables dynamic routing capabilities necessary for complex urban driving. Output modalities traditionally include steering angles and acceleration commands, with emergent techniques increasingly utilizing path-planning outputs like waypoints or cost maps for enhanced interpretability and control precision.

Safety, Interpretability, and Future Directions

The application of end-to-end models in real-world scenarios necessitates robust safety measures and interpretability techniques. The paper examines visual saliency methods for understanding model decision-making processes and auxiliary tasks that enrich model transparency. It further posits the future trajectory of end-to-end systems hinges on resolving interpretability concerns and enhancing the robustness of learning algorithms against adversarial environments. Extensive simulation and comprehensive real-world testing are proposed as essential for validating model efficacy and safety.

In summary, while end-to-end approaches provide a streamlined alternative to modular systems, ongoing research must address the critical challenges of interpretability and safety to facilitate real-world deployment. The survey underscores the prospective advancements and practical implementations anticipated as the field progresses, with implications for AI-driven vehicular autonomy.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (5)
  1. Ardi Tampuu (7 papers)
  2. Maksym Semikin (1 paper)
  3. Naveed Muhammad (5 papers)
  4. Dmytro Fishman (7 papers)
  5. Tambet Matiisen (8 papers)
Citations (198)