As the automotive world moves toward higher levels of driving automation, Level 3 automated driving represents a critical juncture. In Level 3 driving, vehicles can drive alone under limited conditions, but drivers are expected to be ready to take over when the system requests. Assisting the driver to maintain an appropriate level of Situation Awareness (SA) in such contexts becomes a critical task. This position paper explores the potential of Attentive User Interfaces (AUIs) powered by generative Artificial Intelligence ( AI) to address this need. Rather than relying on overt notifications, we argue that AUIs based on novel AI technologies such as large language models or difusion models can be used to improve SA in an unconscious and subtle way without negative efects on drivers overall workload. Accordingly, we propose 5 strategies how generative AIs can be used to improve the quality of takeovers and, ultimately, road safety.
CEUR ceur-ws.org
The advent of automated driving is changing the transportation landscape. The first cars
with Level 3 [
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Current research shows that engagement in non-driving activities, and thus loss of awareness
of the driving environment, can reduce the quality of driver takeovers [
But how can the transition from a state in which the driver can be fully disengaged from the driving task to a state in which the driver must be fully aware of the driving situation to handle a potentially dangerous driving task be made subtly and smoothly? DeGuzman et al. [8] point out that AUIs, that have been shown to efectively manage SA in manual driving, can potentially also be beneficial for automated driving. Other recent work, for example by Wintersberger et al. [9], underlines the potential of AUIs to improve take-over quality. In this position paper we go a step further and argue that in particular the combination of AUIs and generative AI technologies such as Large Language Models (LLMs) and Difusion Models (e.g., Stable Difusion [ 10] or DALLE-3 [11]) can help to subtly bring the driver back into the loop or even subconsciously maintain the required level of SA. When fine-tuned with the rich sensor data available in today’s cars, these models can generate a comprehensive picture of the driving scenario and select guidance strategies tailored to the driving situation and the driver’s state. Not only can they organically guide the driver back to control when the situation requires immediate control, they can also subtly enhance the driver’s SA in situations of increasing uncertainty, where it is not entirely clear whether a take-over will be issued. This prepares the driver without appearing overly cautious.
In the following we present five strategies that employ generative AI and in particular LLMs and Difucion Models to serve as an inspiration for future research.
In the following, we will give a brief overview of current research related to TORs in general and the role that AUIs can play to improve TORs.
In Level 3 automated driving, the automated driving functions can drive the vehicle under
limited conditions [
These findings highlight that for safe takeovers, a holistic understanding of the current driving situation and the state of the driver is important to trigger context-dependent TORs.
Attentive User Interfaces (AUIs) are “computing interfaces that are sensitive to the user’s attention” [21]. These interfaces therefore adapt the type and amount of information displayed based on the attentional state of the user and/or the attentional demands of the environment [8]. For example, due to the driver’s current high stress level and the complex driving situation, an incoming call that’s predicted to be of low urgency, may not be immediately put through, but rather suppressed until the driving situation allows it. Thus, AUIs can not only adjust the timing (e.g., as proposed by Wintersberger et al. [22]) or the visual representation, but also consider the costs and benefits of conflicting actions by taking into account the driver’s state and the driving situation [23].
DeGuzman et al. [8] suggest that AUIs, that have been shown to efectively manage SA in manual driving, may be also beneficial in automated driving. The authors identify several strategies for adapting UIs to either optimize attentional demand or to redirect the driver’s attention to the road. However, they argue that only little research exists that studies the efect of AUIs in automated driving. One of the few studies that show the potential of AUIs for automated driving is presented by Wintersberger et al. [9] who argue that AUIs can improve take-over behavior. Their results show that AUIs improve driving performance, reduce the stress induced to drivers, and reduce the variance in the response times of scheduled TORs.
To efectively tailor the interventions to the driving situation and the driver’s state, an intelligent TOR agent needs access to the driving automation features, the car sensors (e.g., cameras and radar sensors, the cabin cameras) and access to the in-vehicle Human-Machine Interfaces (HMIs) (e.g., infotainment system or head-up display). This information is already available in some modern production cars as shown in the works by Ebel et al. [24, 25]. To personalize interventions, it is also necessary to access personal driver information such as calendar entries. We assume that this information is available by connecting the smartphone to the In-Vehicle Information System (IVIS). Below we present 5 ideas, on how TOR assistants can benefit from generative AI.
Interactive Scenarios Dynamic visual representations of scheduled TORs can improve the usability of TOR assistants [26]. Whereas current research focuses on relatively simple visualizations that are primarily focused on the timing or priority of the TOR, we propose to use generative models such as DALL-E 31 to generate dynamic scenarios that represent the upcoming driving situation. These scenarios can be displayed on the center stack screen as shown in Figure 12, on the head-up display, or on the dashboard. For example, when approaching a highway exit, an image or video sequence of the exit can be displayed, prompting the driver to make a decision. While these scenarios can be used in combination with a direct prompt, they can also be used to subtly prime the driver for an upcoming TOR by displaying dynamic content on the screen in the periphery of the driver’s focus.
Conversational Primers Research suggests that conversational voice assistants and priming techniques can help to build appropriate SA and improve TOR quality [27, 28, 16]. We argue that LLMs can further increase this potential as the system can engage the driver in natural 1https://openai.com/dall-e-3 2Some elements were generated using Adobe Illustrator’s ”Text to Vector Graphic” feature: https://www.adobe.com/ products/illustrator/text-to-vector-graphic.html but brief situation-pendent conversations about the upcoming route or driving scenario. For example, a question such as “Looks like we’re getting of the highway in 10 minutes. Have you driven this route before?” not only informs the driver of the upcoming TOR, but also indirectly prompts the driver to look at the road, thereby improving SA. This strategy can also be useful in situations where the system is uncertain whether a TOR will be triggered in the near future, as the driver may not even realize that the goal of the conversation was to redirect his attention to the road. This way, drivers won’t be annoyed by false positives because they won’t recognize them as such.
Context-Aware and Personalized TORs LLMs can provide concise, contextual descriptions or advice based on real-time sensor data. This information can be used, for example, to generate situation-based TORs: “We are approaching a construction zone on the right lane with a speed limit of 50 km/h, please take control”. While current research suggests that context-aware warnings can lead to safer takeovers [29], these approaches can only detect predefined situations and are therefore limited to specific situations. By leveraging the vast amount of data generated by LLMs and object detection algorithms, TORs are no longer limited to these predefined degrees of freedom. Based on data from the cabin camera, TORs can be tailored not only to the driving situation, but also to the driver’s state and current activity. The intelligent TOR assistant could tell the driver to put away the phone or tablet, arguing that there will be enough time after the construction zone to finish the current activity.
Subtle Nudges Nudging and persuasion can influence drivers to drive more economically [ 30] and more safely [31]. We argue that generative AI technology can be used to generate efective persuasion strategies for TORs. Based on the driver’s past behavior and responses, the generative AI can create tailored priming interventions or use the information gathered from past conversations to persuade the driver to be more aware or take over earlier. For example, the assistant might mention the driver’s daughter’s soccer game to subtly appeal to the driver’s sense of responsibility not to get too distracted.
Ambient Scene Generation Ambient displays and audio cues are an efective measure to improve TOR quality [32, 16]. While current approaches are more or less explicit, we propose that based on the current or upcoming driving situations, an intelligent agent can generate situation-specific ambient scenes. For example, it could subtly change the tone of the infotainment system, or generate soft ambient sounds that resemble the road or trafic to subconsciously focus the driver’s attention on the driving environment. The same applies for ambient lighting. The assistant could gradually synchronize the car’s interior lighting with the outside environment and trafic scene. Dynamic lightning patterns based on passing cars or upcoming situations can be generated and visualized using ambient light technology. A slight change in brightness or hue can alert the driver’s senses without the driver being aware of the change.
Cabin Sensors Vehicle Sensors
Map
V2X
Digital Footprint Interaction
Behavior
Driver State
Estimation Driving Scene Understanding
TOR Generator Digital Persona
Conversation
Agent Scenario Generator
Output IVIS Displays Ambient Light Audio System Tactile Interfaces
We argue that key advantage of using generative AI for scheduled TORs is subtlety and persuasion. The interactions should be smooth, non-intrusive, and feel natural so that the driver’s SA is maintained without the driver actively realizing that they’re being assisted. The goal is not to make the driver dependent on the Intelligent TOR Assistant, but to use the new opportunities that generative AI methods provide to enhance the collaboration between driver and the automated driving system. While subtle cues can help drivers to maintain an appropriate level of SA, LLMs can also be used to generate eloquent and meaningful prompts that persuade the driver to be more attentive. Incorporating personal and situational information could not only improve in-situ TOR quality, but also change driver behavior in the long run.
For all of the strategies presented in this position paper, it is important to emphasize that TORs are safety-critical. Choosing an inappropriate modality or providing false or inaccurate information can have fatal consequences. This needs to be considered in future work, especially in light of current vulnerabilities of generative models such as hallucination, bias, and lack of explainability. In addition, the question of how to ensure that approaches using generative AI methods comply with regulations needs to be answered. Due to their non-deterministic nature, they can’t be evaluated against standardized datasets to assess whether they are “good enough” to be used for safety critical applications3.
While some of the above strategies may seem dystopian at the time of this writing, a digital assistant that is intimately aware of user preferences and behaviors and can carry on a conversation as naturally as a human counterpart may be technically possible and socially acceptable in just a few years. However, research suggests that conversational agents that seem too human don’t necessarily drive adoption. In fact, they may deter people from using the technology [33]. Thus, implementing strategies such as the Subtle Nudges strategy is a challenging endeavor and more research is needed to enable systems such as the one presented in this position paper. 3Not to say that the question of what is “good enough” when it comes to automated driving has been answered yet. [5] P. Marti, C. Jallais, A. Koustanaï, A. Guillaume, F. Mars, Impact of the driver’s visual engagement on situation awareness and takeover quality, Transportation Research Part F: Trafic Psychology and Behaviour 87 (2022) 391–402. doi: 10.1016/j.trf.2022.04.018. [6] S. Ma, W. Zhang, Z. Yang, C. Kang, C. Wu, C. Chai, J. Shi, Y. Zeng, H. Li, Take over Gradually in Conditional Automated Driving: The Efect of Two-stage Warning Systems on Situation Awareness, Driving Stress, Takeover Performance, and Acceptance, International Journal of Human–Computer Interaction 37 (2021) 352–362. doi:10.1080/10447318.2020.1860514. [7] S. Agrawal, S. Peeta, Evaluating the impacts of situational awareness and mental stress on takeover performance under conditional automation, Transportation Research Part F: Trafic Psychology and Behaviour 83 (2021) 210–225. doi: 10.1016/j.trf.2021.10.002. [8] C. A. DeGuzman, D. Kanaan, B. Donmez, Attentive User Interfaces: Adaptive Interfaces that Monitor and Manage Driver Attention, in: A. Riener, M. Jeon, I. Alvarez (Eds.), User Experience Design in the Era of Automated Driving, volume 980, Springer International Publishing, Cham, 2022, pp. 305–334. doi:10.1007/978-3-030-77726-5_12. [9] P. Wintersberger, C. Schartmüller, A. Riener, Attentive User Interfaces to Improve Multitasking and Take-Over Performance in Automated Driving: The Auto-Net of Things, International Journal of Mobile Human Computer Interaction 11 (2019) 40–58. doi:10.4018/IJMHCI.2019070103. [10] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High-Resolution Image Synthesis with Latent Difusion Models, in: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, New Orleans, LA, USA, 2022, pp. 10674–10685. doi:10. 1109/CVPR52688.2022.01042. [11] J. Betker, G. Goh, L. Jing, †. TimBrooks, J. Wang, L. Li, †. LongOuyang, †. JuntangZhuang, †. JoyceLee, †. YufeiGuo, †. WesamManassra, †. PrafullaDhariwal, †. CaseyChu, †. YunxinJiao, A. Ramesh, Improving image generation with better captions, 2023. [12] W. Morales-Alvarez, O. Sipele, R. Léberon, H. H. Tadjine, C. Olaverri-Monreal, Automated Driving: A Literature Review of the Take over Request in Conditional Automation, Electronics 9 (2020) 2087. doi:10.3390/electronics9122087. [13] P. Wintersberger, P. Green, A. Riener, Am I Driving or Are You or Are We Both? A Taxonomy for Handover and Handback in Automated Driving, in: Proceedings of the 9th International Driving Symposium on Human Factors in Driver Assessment, Training, and Vehicle Design: Driving Assessment 2017, University of Iowa, Manchester Village, Vermont, USA, 2017, pp. 333–339. doi:10.17077/drivingassessment.1655. [14] A. Eriksson, N. A. Stanton, Takeover Time in Highly Automated Vehicles: Noncritical Transitions to and From Manual Control, Human Factors: The Journal of the Human Factors and Ergonomics Society 59 (2017) 689–705. doi:10.1177/0018720816685832. [15] R. McCall, F. McGee, A. Mirnig, A. Meschtscherjakov, N. Louveton, T. Engel, M. Tscheligi, A taxonomy of autonomous vehicle handover situations, Transportation Research Part A: Policy and Practice 124 (2019) 507–522. doi:10.1016/j.tra.2018.05.005. [16] S. Sadeghian Borojeni, L. Weber, W. Heuten, S. Boll, From reading to driving: Priming mobile users for take-over situations in highly automated driving, in: Proceedings of the 20th International Conference on Human-Computer Interaction with Mobile Devices and Services, ACM, Barcelona Spain, 2018, pp. 1–12. doi:10.1145/3229434.3229464. [17] J. Radlmayr, C. Gold, L. Lorenz, M. Farid, K. Bengler, How Trafic Situations and NonDriving Related Tasks Afect the Take-Over Quality in Highly Automated Driving, Proceedings of the Human Factors and Ergonomics Society Annual Meeting 58 (2014) 2063–2067. doi:10.1177/1541931214581434. [18] C. Gold, D. Damböck, L. Lorenz, K. Bengler, “Take over!” How long does it take to get the driver back into the loop?, Proceedings of the Human Factors and Ergonomics Society Annual Meeting 57 (2013) 1938–1942. doi:10.1177/1541931213571433. [19] A. Feldhutter, A. Ruhl, A. Feierle, K. Bengler, The Efect of Fatigue on Take-over Performance in Urgent Situations in Conditionally Automated Driving, in: 2019 IEEE Intelligent Transportation Systems Conference (ITSC), IEEE, Auckland, New Zealand, 2019, pp. 1889–1894. doi:10.1109/ITSC.2019.8917183. [20] S. H. Yoon, Y. W. Kim, Y. G. Ji, The efects of takeover request modalities on highly automated car control transitions, Accident Analysis & Prevention 123 (2019) 150–158. doi:10.1016/j.aap.2018.11.018. [21] R. Vertegaal, Attentive User Interfaces, Communications of the ACM 46 (2003) 30–33.
doi:10.1145/636772.636794. [22] P. Wintersberger, A. Riener, C. Schartmüller, A.-K. Frison, K. Weigl, Let Me Finish before I Take Over: Towards Attention Aware Device Integration in Highly Automated Vehicles, in: Proceedings of the 10th International Conference on Automotive User Interfaces and Interactive Vehicular Applications, ACM, Toronto ON Canada, 2018, pp. 53–65. doi:10. 1145/3239060.3239085. [23] M. Braun, F. Weber, F. Alt, Afective Automotive User Interfaces–Reviewing the State of Driver Afect Research and Emotion Regulation in the Car, ACM Computing Surveys 54 (2022) 1–26. doi:10.1145/3460938. [24] P. Ebel, C. Lingenfelder, A. Vogelsang, On the forces of driver distraction: Explainable predictions for the visual demand of in-vehicle touchscreen interactions, Accident Analysis & Prevention 183 (2023) 106956. doi:10.1016/j.aap.2023.106956. [25] P. Ebel, K. J. Gülle, C. Lingenfelder, A. Vogelsang, Exploring Millions of User Interactions with ICEBOAT: Big Data Analytics for Automotive User Interfaces, in: AutomotiveUI ’23: 15th International Conference on Automotive User Interfaces and Interactive Vehicular Applications, ACM, Ingolstadt, Germany, 2023. doi:10.48550/arXiv.2307.06089. [26] K. Holländer, B. Pfleging, Preparing Drivers for Planned Control Transitions in Automated Cars, in: Proceedings of the 17th International Conference on Mobile and Ubiquitous Multimedia, ACM, Cairo Egypt, 2018, pp. 83–92. doi:10.1145/3282894.3282928. [27] K. Mahajan, D. R. Large, G. Burnett, N. R. Velaga, Exploring the benefits of conversing with a digital voice assistant during automated driving: A parametric duration model of takeover time, Transportation Research Part F: Trafic Psychology and Behaviour 80 (2021) 104–126. doi:10.1016/j.trf.2021.03.012. [28] X. Bai, J. Feng, Unlocking Safer Driving: How Answering Questions Help Takeovers in Partially Automated Driving, Proceedings of the Human Factors and Ergonomics Society Annual Meeting (2023) 21695067231192202. doi:10.1177/21695067231192202. [29] E. Pakdamanian, E. Hu, S. Sheng, S. Kraus, S. Heo, L. Feng, Enjoy the Ride Consciously with CAWA: Context-Aware Advisory Warnings for Automated Driving, in: Proceedings of the 14th International Conference on Automotive User Interfaces and Interactive Vehicular Applications, ACM, Seoul Republic of Korea, 2022, pp. 75–85. doi:10.1145/3543174.3546835. [30] A. Meschtscherjakov, D. Wilfinger, T. Scherndl, M. Tscheligi, Acceptance of future persuasive in-car interfaces towards a more economic driving behaviour, in: Proceedings of the 1st International Conference on Automotive User Interfaces and Interactive Vehicular Applications, ACM, Essen Germany, 2009, pp. 81–88. doi:10.1145/1620509.1620526. [31] V. Choudhary, M. Shunko, S. Netessine, S. Koo, Nudging Drivers to Safety: Evidence from a Field Experiment, Management Science 68 (2022) 4196–4214. doi:10.1287/mnsc.2021. 4063. [32] S. Sadeghian Borojeni, L. Chuang, W. Heuten, S. Boll, Assisting Drivers with Ambient Take-Over Requests in Highly Automated Driving, in: Proceedings of the 8th International Conference on Automotive User Interfaces and Interactive Vehicular Applications, ACM, Ann Arbor MI USA, 2016, pp. 237–244. doi:10.1145/3003715.3005409. [33] T. Fernandes, E. Oliveira, Understanding consumers’ acceptance of automated technologies in service encounters: Drivers of digital voice assistants adoption, Journal of Business Research 122 (2021) 180–191. doi:10.1016/j.jbusres.2020.08.058.