Learning Accurate, Comfortable and Human-like Driving
DOI: 10.48550/arxiv.1903.10995
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 addresses the challenge of developing autonomous driving systems that are not only accurate but also comfortable and human-like, arguing that these qualities are essential for public acceptance and safe interaction with human drivers. While previous research has primarily focused on driving accuracy, this work formalizes a unified learning framework that simultaneously optimizes for accuracy, passenger comfort, and human-like behavior. The authors identify three key gaps in existing end-to-end driving models: the underutilization of high-fidelity numerical map data, the lack of explicit comfort measures in learning algorithms, and the absence of mechanisms to ensure driving styles resemble human behavior. To address these issues, the authors propose a sequence-based end-to-end learning model trained on the Drive360 dataset, which contains 60 hours and 3,000 km of real-world driving data. The method incorporates three main innovations. First, it augments visual inputs with engineered numerical features from HERE Technologies maps, such as distances to intersections, speed limits, and road curvature, processed via LSTM and fully connected networks. Second, it improves the learning procedure from pointwise to sequence-based prediction, embedding a comfort loss that minimizes longitudinal and lateral jerk (the second derivative of steering and speed) to reduce motion sickness. Third, it employs adversarial learning to achieve human-like driving; a discriminator network is trained to distinguish between human and machine driving maneuvers, while the driving model (generator) is trained to fool this discriminator, thereby aligning its behavior with human patterns. Experiments demonstrate that the proposed model outperforms previous state-of-the-art methods, including those using only visual maps or no map data. The integration of HERE numerical features significantly improves driving accuracy, particularly in challenging scenarios like approaching intersections and traffic lights, reducing steering angle error from 19.69 to 16.83. The inclusion of the comfort loss substantially reduces lateral and longitudinal oscillations, enhancing ride comfort with only a modest trade-off in accuracy. Furthermore, the adversarial component successfully makes the vehicle’s driving style more indistinguishable from human drivers. The study concludes that combining numerical map data, sequence-based comfort optimization, and adversarial human-likeness training results in a driving model that is more accurate, comfortable, and human-like than prior approaches, offering a more robust foundation for autonomous vehicle deployment.
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.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | OpenAlex-citations | — | — | 1 | 2026-06-25 |
| archive | success | openalex | — | — | 5 | 2026-06-26 |
| extract | success | cached | — | — | 2 | 2026-06-26 |
| clean | success | clean | — | — | 1 | 2026-06-25 |
| chunk | success | chunk | — | — | 1 | 2026-06-25 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-25 |
| promote | success | — | — | — | 1 | 2026-06-25 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 1 | 2026-06-26 |
| tag | success | vector_similarity | — | — | 6 | 2026-06-25 |
| verify | success | — | — | — | 1 | 2026-06-26 |
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.
Information type
What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).
- Methodological Resource: dataset resource, tool software
- Theoretical Contribution: computational model