DISC : Dataset for Analyzing D riving Styles I n S imulated C rashes for Mixed Autonomy
Sandip Sharan Senthil Kumar, Sandeep Thalapanane · 2025
Read DISC : Dataset for Analyzing D riving Styles I n S imulated C rashes for Mixed Autonomy by Sandip Sharan Senthil Kumar, Sandeep Thalapanane (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-00050.txt.
Introduction
Sandip Sharan Senthil Kumar and co-authors introduce a data set on how human drivers act just before crashes, in a paper from 2025. Self-driving cars still struggle with the seconds before a crash. Real data on those moments are rare, and mixed traffic with human drivers adds to the challenge.
The data set, called DISC, was gathered with a virtual reality driving simulator built by the team. Hundreds of drivers met twelve staged crash scenarios in a virtual city. Their driving styles were sorted into more than eight classes, and surveys added standard traits for each driver. The data support tools that classify pre-crash behavior and predict paths for single drivers. The goal is safer responses by self-driving cars near human drivers.
Catalog record
| CREATOR | Sandip Sharan Senthil Kumar, Sandeep Thalapanane |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 32 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-00050
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
DISC : Dataset for Analyzing D riving Styles I n S imulated C rashes for Mixed Autonomy Thanks: The authors are with the Department of Computer Science, University of Maryland at College Park, MD, U.S.A. E-mail: {sandeept, sandip26, guruadp, sourang, lyzheng, lin}@umd.edu Sandip Sharan Senthil Kumar, Sandeep Thalapanane Abstract Handling pre-crash scenarios is still a major challenge for self-driving cars due to limited practical data and human-driving behavior datasets. We introduce DISC (Driving Styles In Simulated Crashes), one of the first datasets designed to capture various driving styles and behaviors in pre-crash scenarios for mixed autonomy analysis. DISC includes over 8 classes of driving styles/behaviors from hundreds of drivers navigating a simulated vehicle through a virtual city, encountering rare-event traffic scenarios. This dataset enables the classification of pre-crash human driving behaviors in unsafe conditions, supporting individualized trajectory prediction based on observed driving patterns. By utilizing a custom-designed VR-based in-house driving simulator, TRAVERSE , data was collected through a driver-centric study involving human drivers encountering twelve simulated accident scenarios. This dataset fills a critical gap in human-centric driving data for rare events involving interactions with autonomous vehicles. It enables autonomous systems to better react to human drivers and optimize trajectory prediction in mixed autonomy environments involving both human-driven and self-driving cars. In addition, individual driving behaviors are classified through a set of standardized questionnaires, carefully designed to identify and categorize driving behavior traits. We correlate data features with driving behaviors, showing that the simulated environment reflects real-world driving styles. DISC is the first dataset to capture how various driving styles respond to accident scenarios, offering significant potential to enhance autonomous vehicle safety and driving behavior analysis in mixed autonomy environments. Fig. 1: Scenarios from the driver’s point of view: The LEFT image depicts the deer-crossing scenario, where the participant encounters a running deer while driving on a country road. The MIDDLE image shows the running red lights scenario, where a participant encounters a car at an intersection that crosses the road after ignoring the red traffic signal. The RIGHT image displays a jaywalking pedestrian from the driver’s point of view. I INTRODUCTION Recent advances in self-driving platforms, from food delivery robots to autonomous vehicles (AV), highlight the urgent need for strong safety standards. Leading companies in the autonomous vehicle industry have pioneered self-driving taxis capable of navigating complex traffic situations, often achieving safety levels surpassing those of human drivers. Despite these advancements, real-world data collection, using sensors such as LiDAR and cameras, is often limited to mostly normal driving scenarios and particularly lacking in accidental events. This paper addresses a considerable gap in current research: the lack of adequate data collected on rare events and accidents . In order for autonomous vehicles (AVs) to learn to handle adversarial events (e.g. accidents), compounded by unpredictable factors such as jaywalking pedestrians and adverse weather conditions, it remains a major challenge to guarantee safe autonomous driving. Current autonomous driving systems are predominantly trained using datasets collected from safe, typical traffic conditions. While this approach ensures robustness for normal driving scenarios, it inadvertently leaves AVs ill-prepared for unexpected and dangerous scenarios. This scarcity of data on adverse driving events results in AVs often not performing optimally when encountered with rare and dangerous scenarios. Addressing this issue requires innovative methods for generating realistic and high-quality training data encompassing a broader spectrum of unsafe and hazardous driving conditions. Typically, synthetic risky scenarios are created using simulations, but these simulations frequently lack the realism and accuracy needed to fully capture the complexity of adversarial events. Furthermore, in the near horizon during the deployment of self-driving vehicles, we can expect the AVs to operate in a ‘ mixed autonomy ’, where there will be human-driven and self-driving cars operating side by side on the road for the foreseeable future. AVs require the capability to anticipate human-driven vehicle behaviors that will likely be different and less predictable than a self-driving car, thereby making it especially critical to capture large amounts of human driving data under diverse pre-crash scenarios. This work aims to bridge this gap by employing a virtual reality (VR) vehicle simulator, TRAVERSE [ 1 ] , to immerse participants in various accident scenarios. This approach enables safe collections of human driving data and behaviors in high-risk situations, significantly enhancing our understanding of driving dynamics and safety. TRAVERSE is specifically designed to simulate rare event scenarios based on the National Highway Traffic Safety Administration (NHTSA) pre-crash typology [ 2 ] , ensuring that the assessed scenarios are both impactful in the real world via quantified societal costs, in addition to being realistic and relevant for participants. Participants also complete a driving style test based on the Multidimensional Driving Style Inventory (MDSI) [ 3 ] questionnaire to analyze their driving style. This comprehensive questionnaire enables us to classify driving behaviors into eight distinct traits as identified by the MDSI, providing valuable insights into driving styles. This paper presents three key contributions: • The DISC dataset comprises 2,527,004 sets of sensory data for 1,205 vehicle trajectories from 110 participants, capturing diverse driving styles and behaviors in pre-crash scenarios for mixed autonomy analysis, providing a critical foundation for trajectory prediction models. • Each collected set of sensory data consists of braking, acceleration, steering, and eye-tracking metrics, collected using Meta Quest Pro, which allows for a deeper analysis of driver perception and responses. • DISC consists of diverse data generated under adverse weathers (fog, snow, rain) and varying time/lighting settings (dawn, morning, noon, dusk, night). In summary, our approach utilizes VR-based simulations to produce essential training data for autonomous vehicles in high-risk scenarios. By replicating real-world traffic conditions, we enhance the accuracy of models predicting human-driven vehicle behaviors, thereby improving autonomous vehicle navigation in mixed-autonomy environments. By leveraging the newly introduced human-driving style dataset, DISC , we demonstrate significant potential for advancing the safety and efficiency of autonomous driving systems. II Related Works II-A Driving Simulators High-fidelity driving simulators, such as the NADS1 Driving Simulator [ 4 , 5 , 6 ] and the WTI Simulator [ 7 ] , offer exceptional realism but face deployment challenges due to their high costs and logistical requirements. In contrast, low-fidelity simulators present more cost-effective and portable options. The University of Iowa’s miniSim [ 8 ] strikes a balance between realism and practicality, while Deep Drive [ 9 ] , is tailored for self-driving AI, utilizing Unreal Engine. CARLA [ 10 ] provides detailed urban environments with dynamic traffic and weather, and LGSVL [ 11 ] enhances these features with physics-based models and Unity compatibility. However, these simulators often fall short of realism which is vital for behavioral studies. To address this gap, VR simulators such as DrEyeVR [ 12 ] , built on CARLA, offer improved realism but face challenges with customizability and frame rates. Our in-house simulator, TRAVERSE, integrates Unity, SUMO, and Road Runner [ 13 ] to create dynamic driving scenarios with environmental effects, enhancing the realism for behavioral research. II-B Driving Datasets II-B 1 Sensory and Motion Forecasting Datasets Recent advancements in autonomous driving datasets have significantly improved scenario realism. The Waymo Open Motion Dataset [ 14 ] offers extensive data on perception and motion forecasting in diverse urban settings. The Argoverse-2