Great GATsBi: Hybrid, Multimodal, Trajectory Forecasting for Bicycles using Anticipation Mechanism
Kevin Riehl, Shaimaa K. El-Baklish, Anastasios Kouvelas, Michail A. Makridis · 2025
Read Great GATsBi: Hybrid, Multimodal, Trajectory Forecasting for Bicycles using Anticipation Mechanism by Kevin Riehl, Shaimaa K. El-Baklish, Anastasios Kouvelas, Michail A. Makridis (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-2508-14523.txt.
Introduction
Riehl, El-Baklish, Kouvelas, and Makridis present Great GATsBi, a system that forecasts where bicycles will move next. Most past work focused on walkers and motor vehicles, yet bike riders account for many road deaths. Their 2025 framework blends physics based and social based views of bike motion.
Physics models capture how a bike moves like a small vehicle, and they do well over short horizons. A graph based network models social links with nearby road users, including guesses about their near future paths, and does well over longer horizons. Blending both beat prior top methods in their tests. A group cycling experiment further checked the system in a live setting.
Catalog record
| CREATOR | Kevin Riehl, Shaimaa K. El-Baklish, Anastasios Kouvelas, Michail A. Makridis |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 62 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-14523
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Great GATsBi: Hybrid, Multimodal, Trajectory Forecasting for Bicycles using Anticipation Mechanism Kevin Riehl, Shaimaa K. El-Baklish, Anastasios Kouvelas, Michail A. Makridis Abstract Accurate prediction of road user movement is increasingly required by many applications ranging from advanced driver assistance systems to autonomous driving, and especially crucial for road safety. Even though most traffic accident fatalities account to bicycles, they have received little attention, as previous work focused mainly on pedestrians and motorized vehicles. In this work, we present the Great GATsBi , a domain-knowledge-based, hybrid, multimodal trajectory prediction framework for bicycles. The model incorporates both physics-based modeling (inspired by motorized vehicles) and social-based modeling (inspired by pedestrian movements) to explicitly account for the dual nature of bicycle movement. The social interactions are modeled with a graph attention network, and include decayed historical, but also anticipated, future trajectory data of a bicycles neighborhood, following recent insights from psychological and social studies. The results indicate that the proposed ensemble of physics models – performing well in the short-term predictions – and social models – performing well in the long-term predictions – exceeds state-of-the-art performance. We also conducted a controlled mass-cycling experiment to demonstrate the framework’s performance when forecasting bicycle trajectories and modeling social interactions with road users. 1 Introduction Accurate prediction of road user movement is crucial for various applications, including urban planning, infrastructure design, and intelligent transportation systems. Effective tracking and surveillance of road users can provide timely information, enabling smart signalized intersections yu2023v2x . Moreover, precise forecasts of neighboring road users enhance safety systems in advanced driver assistance technologies, helping prevent collisions and accidents kosaraju2019social . Autonomous vehicles, in particular, frequently interact with neighboring road users and require accurate perceptions of both current and future positions to navigate safely madjid2025trajectory . Trends in urbanization, municipal sustainability efforts, increasing issues with congestion and emissions, and technological innovations drive the increasing use of active modes, like e-bikes and bicycles. Each year, road traffic accidents result in approximately 1.24 million fatalities, with vulnerable road users (VRUs) and especially cyclists comprising the largest proportion of these deaths toroyan2013launches ; gao2021trajectory ; zernetsch2016trajectory ; bessi2023trajectory . Despite the increasing trend in severe traffic accidents involving bicycles and VRUs de2020injuries ; juhra2012bicycle , their trajectory forecasting remains heavily underrepresented in the literature, that primarily focuses on pedestrians and motorized vehicles huang2022survey ; ding2023incorporating ; schuetz2023review ; madjid2025trajectory . Bicycle trajectory prediction is particularly challenging due to their hybrid behavioral characteristics. Bicycles and motorized vehicles alike, are subject to non-holonomic kinematic constraints and can reach relatively high speeds. Similarly, maneuver-related planning patterns such as following, lane changing and overtaking can be observed gao2021trajectory . However, unlike cars, bicycles are often less constrained by lane boundaries and can exhibit more flexible and sudden-change behaviors (in direction), similar to pedestrians arlauskas2025 ; li2023two ; brunner2024microscopic . This combination of vehicle-like dynamics and pedestrian-like maneuverability complicates accurate trajectory prediction, especially in mixed-traffic environments li2023two . Previous works on bicycle trajectory forecasting primarily modeled bicycle behavior similar to pedestrians. By focusing on intent pool2019context ; gao2021trajectory , sudden change behavior li2023two and multiple interactions with the neighborhood huang2021lstm ; li2023two when modeling social context features, they neglect physical, kinematics-related properties of bicycles. Moreover, previous efforts are characterized by data-driven approaches only, overlooking the potential of incorporating domain-knowledge with learning, and impeding interpretability borghesi2020improving ; dash2022review . To address the challenges of bicycle trajectory forecasting and related limitations, we propose the great GATsBi , a hybrid, domain knowledge-based framework, that leverages both social and physical modeling, for multimodal trajectory forecasting of bicycles. First, we model the social context as a graph structure using graph attention networks (GATs) and explicitly incorporate anticipation and perception decay following recent insights from behavioral modeling hastie2022schematic ; tarder2024brain ; kinsky2020trajectory ; tanke2023social ; tanke2023social ; ruan2024learning . Second, we combine kinematic and extended Kalman-filtering forecasting models to a deep ensemble embedding for the physics context. Third, we follow a hybrid approach, that combines these two, domain-knowledge-based contextual features, to gain improvements in forecasting and enabling insights for interpretability. Doing so, we account for the dual nature of bicycles, that share behavioral aspects with pedestrians and motorized vehicles. We conducted a real-world, closed-loop, mass cycling experiment that enabled the observation of various interactions and bicycle maneuvers in varying traffic density contexts without the interference of road-contextual properties. A comparative benchmark on the trajectory dataset supports the contributions of the proposed method when compared with common forecasting methods for pedestrians and vehicles. The trajectory dataset, and implementation will be made publicly accessible in an open-source repository upon publication. \includegraphics [width=1.0]Figure_Architecture_1.pdf Figure 1: Problem Statement: Trajectory Prediction. This work is concerned with forecasting the trajectories of moving road users given their neighbors and their own (ego) historical trajectory information. 2 Related Work Trajectory forecasting typically involves contextual factors related to physics, road conditions, and social interactions. Prediction models can be categorized into physics-based methods, such a Kalman filtering approaches, and machine learning-based methods, which are better suited for capturing road and social context. These models can predict various types of outcomes, including a single trajectory (unimodal), a probabilistic distribution of possible future trajectories (multimodal), and driver intentions (maneuvers) huang2022survey . Recent works especially leveraged the potential of machine and deep learning (ML) based approaches ding2023incorporating . Bicylce Trajectory Forecasting. Contrary to previous work that relied on kinematic modeling approaches, zernetsch2016trajectory combined a simple physics model with a multi-layer perceptron (MLP) to achieve better predictions. pool2017using showed that combining road topology features with the bicycle’s historical trajectory further enhances forecasts. pool2019context ; gao2021trajectory highlighted the importance of learning the latent intention of the cyclist, which allows the model to leverage behavioral cues in trajectory prediction. li2023two expanded on this by differentiating between intentions and latent sudden change behavior to better account for behavioral heterogeneity. huang2021lstm demonstrated that incorporating neighborhood information to model multiple interactions between the cyclist and other road users can yield further improvements. Besides, a growing branch of work explored fast-inference, learning-based sensor fusion of onboard sensors for equipped, smart bicycles koornstra2023predicting ; arlauskas2025 ; sass2023autonomous ; bessi2023trajectory . Despite these advances, the former works are limited in three aspects. First, they neglected the multimodal nature of bicycle behavior which was shown to significantly improve forecasting accuracy in a pedestrian context kosaraju2019social . Second, they focused on social context and overlooked the importance of systematic combination of physics-knowledge-based models. Third, their neighborhood was modeled over-simplistically in the form of focal attention mechanisms. To address these gaps, we propose a multimodal forecasting model based on Gaussian Mixture Models (GMM), explicitly combining physics-knowledge and social-knowledge to improve forecasting accuracy, and model the bicycle’s social context as a graph using graph attention networks (GATs). Human Motion Modeling. Traditionally, human motion modeling was often characterized by analytical modeling approaches. Notably, Helbing’s Social Force Model helbing1995social explicitly takes surrounding individuals and human-human interactions into account when modeling human motion. Modern approaches yield significantly improved forecasts leveraging ML approaches, such as SocialLSTM alahi2016social . Social-BiGAT kosaraju2019social highlighted the importance of multimodality in the context of motion modeling, and explicitly reflected the social context in a graph neural network (GNN). In our model, we draw upon these advancements from pedestrian modeling and transfer them into the