Discomfort detection during automated driving using temporal transformers

Assarzadeh, Maha; Hartwich, Franziska; Vitay, Julien; Bocklisch, Franziska; Hamker, Fred H. · 2025 · Crossref

DOI: 10.3389/fcomp.2025.1639505

archive: archived pipeline: cataloged verified

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

Summary

This study addresses the challenge of detecting passenger discomfort during automated driving, a critical factor for user acceptance and trust in autonomous vehicles. The authors frame discomfort as a time-varying psychological state influenced by passenger physiology, environmental conditions, and vehicle automation behaviors. While previous research has utilized static machine learning models or simple recurrent networks, this paper introduces the Temporal Fusion Transformer (TFT), an advanced attention-based deep learning architecture, to dynamically predict discomfort levels. The TFT was selected for its ability to handle long-term temporal dependencies, integrate heterogeneous sensor inputs, and provide interpretable insights into feature importance, which are essential for understanding human-technology interaction. The research utilizes a dataset from a simulated driving experiment involving 100 first-time users of fully automated driving (SAE Level 5). Participants experienced a standardized 7 km test track comprising urban and rural scenarios with varying complexity. Discomfort was measured continuously via a manual handset control (scale 0–100) at 60 Hz. Input features included physiological data (heart rate, pupil diameter, blink rate) recorded via smart bands and eye-tracking glasses, as well as environmental and vehicle state data from the simulator. The authors developed two TFT models: TFT-full, using all available features, and TFT-restricted, using a subset of inputs. These were compared against DeepAR, an autoregressive baseline model. The study systematically analyzed the impact of input window sizes on prediction performance, ensuring the model predicts future discomfort based solely on past sensor signals without using past discomfort values to prevent information leakage. The results demonstrate that the TFT-restricted model achieved superior performance compared to both TFT-full and DeepAR. Specifically, with a window size of 300 time steps (approximately 5 seconds), the TFT-restricted model yielded a mean absolute error (MAE) of 0.037 and a root mean square error (RMSE) of 0.131. The study confirms that the TFT effectively captures temporal dependencies in the data, allowing for accurate multi-horizon forecasting of discomfort levels. The restricted model’s success suggests that a carefully selected subset of physiological and environmental features is sufficient for high-accuracy prediction, potentially reducing computational overhead in real-world applications. The significance of this work lies in its demonstration that transformer-based architectures can effectively model the complex, time-dependent nature of passenger discomfort. By providing interpretable explanations of which features drive predictions, the TFT facilitates the development of adaptive automated driving systems that can trigger counteracting measures, such as adjusting driving styles or in-vehicle information, to enhance passenger comfort. This approach supports the broader goal of improving human-technology cooperation and trust in automated vehicles, although the authors note that scalability to real-world environments remains a subject for future research.

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.

StageOutcomeToolModelPromptAttemptsCompleted
discover success Crossref 1 2026-06-05
archive success canonical_url 1 2026-06-25
extract success cached 2 2026-06-25
clean success clean 1 2026-06-07
chunk success chunk 1 2026-06-07
embed success embed Qwen/Qwen3-Embedding-8B 1 2026-06-07
promote success 1 2026-06-05
summarize success llm qwen3.6-27b-prismaquant summ-v5 1 2026-06-25
tag success vector_similarity 15 2026-06-11
verify success 1 2026-06-26

Summary generated by qwen3.6-27b-prismaquant on 2026-06-25; 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).