Autonomous vehicles (AVs) are set to transform transportation by providing safer, more efficient, and accessible mobility solutions. However, deploying AV systems requires designers to ensure these vehicles can navigate complex, dynamic traffic environments safely and precisely. A vital component of this capability is the ability to predict the behavior of surrounding road users, yet achieving reliable predictions is a complex task. This thesis investigates several challenges in trajectory and intention prediction for autonomous driving, focusing on contextual awareness, probabilistic modeling, and differential motion constraints.

A primary focus of this thesis is context awareness, which includes interaction-aware (agent-to-agent) and environment-aware (road-to-agent) modeling. Early approaches to context awareness involved manually crafting interaction features. While effective, these methods rely on predefined heuristics and often scale poorly as environmental complexity increases. To address these limitations, the thesis adopts a graph-based approach that offers greater flexibility and expressiveness. By constructing relational graphs, graph neural networks can be used to learn agent interactions and environmental context in a data-driven manner. The thesis proposes several context-aware models and provides an extensive evaluation of their mechanisms, highlighting their overall impact on prediction performance.

Another core theme of this thesis is addressing the inherent uncertainty and non-determinism of traffic environments. This involves creating models that provide probabilistic predictions. Given the multimodal nature of road-traffic agent behavior, it is also important to design methods that offer multiple candidate predictions for a single condition, enabling AVs to account for different possible future outcomes. This thesis proposes several context-aware frameworks that leverage probabilistic modeling, illustrating how ensemble methods, mixture density networks, and diffusion-based generative models can be adapted to provide uncertainty estimates and multimodal predictions.

While neural networks are well-suited for capturing the complex dynamics of traffic, they do not inherently ensure that outputs conform to physical laws, which poses risks in safety-critical applications. To address this, the thesis incorporates differential motion constraints into the prediction framework to ensure that predicted trajectories are accurate, physically feasible, and robust to noise. In addition to improving prediction accuracy, it is demonstrated how this integration enhances interpretation, extrapolation, and generalization capabilities.

The use of neural ordinary differential equations is a central component of this thesis, providing a data-driven approach to modeling the motion of agents—such as pedestrians—that are challenging to describe using physical laws or rule-based methods. The thesis investigates their application beyond motion prediction to a wide array of sequence modeling tasks, analyzing the effects of numerical integration techniques, stability regions, and initialization methods on model performance. A key contribution is the proposal of a stability-informed initialization (SII) technique, which significantly enhances model convergence, training stability, and prediction accuracy across various learning benchmarks.

This paper addresses the training of Neural Ordinary Differential Equations (neural ODEs), and in particular explores the interplay between numerical integration techniques, stability regions, step size, and initialization techniques. It is shown how the choice of integration technique implicitly regularizes the learned model, and how the solver’s corresponding stability region affects training and prediction performance. From this analysis, a stability-informed parameter initialization technique is introduced. The effectiveness of the initialization method is displayed across several learning benchmarks and industrial applications.

Future progress toward the realization of fully self-driving vehicles still re-quires human-level social compliance, arguably dependent on the ability to accurately forecast the behavior of surrounding road users. Due to the inter-connected nature of traffic participants, in which the actions of one agent can significantly influence the decisions of others, the development of behavior pre-diction methods is crucial for achieving resilient autonomous motion planning. As high-quality data sets become more widely available and many vehicles already possess significant computing power, the possibility of adopting a data-driven approach for motion prediction is increasing. 

The first contribution is the design of an intention-prediction model based on autoencoders for highway scenarios. Specifically, the method targets the problem of data imbalance in highway traffic data using ensemble methods and data-sampling techniques. The study shows that commonly disregarded information holds potential use for improved prediction performance and the importance of dealing with the data imbalance problem. 

The second contribution is the development of a probabilistic motion pre-diction framework. The framework is used to evaluate various graph neural network architectures for multi-agent prediction across various traffic scenarios. The graph neural network computes the inputs to an underlying motion model, parameterized using neural ordinary differential equations. The method additionally introduces a novel uncertainty propagation approach by combining Gaussian mixture modeling and extended Kalman filtering techniques. 

The third contribution is attributed to the investigation of combing data-driven models with motion modeling and methods for numerical integration. The study illustrates that improved prediction performance can be achieved by the inclusion of differential constraints in the model, but that the choice of motion model as well as numerical solver can have a large impact on the prediction performance. It is also shown that the added differential constraints improve extrapolation properties compared to complete black-box approaches. 

The thesis illustrates the potential of data-driven methods and their usability for the behavior prediction problem. Still, there are additional challenges and interesting questions to investigate—the main one being the investigation of their use in autonomous navigation frameworks. 

Enabling resilient autonomous motion planning requires robust predictions of surrounding road users’ future behavior. In response to this need and the associated challenges, we introduce our model titled MTP-GO. The model encodes the scene using temporal graph neural networks to produce the inputs to an underlying motion model. The motion model is implemented using neural ordinary differential equations where the state-transition functions are learned with the rest of the model. Multimodal probabilistic predictions are obtained by combining the concept of mixture density networks and Kalman filtering. The results illustrate the predictive capabilities of the proposed model across various data sets, outperforming several state-of-the-art methods on a number of metrics.

To study the effects of uncertainty in autonomous motion planning and control, an 8-DOF model of a tractor-semitrailer is implemented and analyzed. The implications of uncertainties in the model are then quantified and presented using sensitivity analysis and closed-loop simulations. The study shows that different model parameters are more or less critical depending on the investigated scenario.- Using sampling-based closed-loop predictions, uncertainty bounds on state variable trajectories are determined. Our findings suggest the potential for the inclusion of our method within a robust predictive controller or as a driver-assistance system for rollover or lane departure warning.

The use of learning-based methods for vehicle behavior prediction is a promising research topic. However, many publicly available data sets suffer from class distribution skews which limits learning performance if not addressed. This paper proposes an interaction-aware prediction model consisting of an LSTM autoencoder and SVM classifier. Additionally, an imbalanced learning technique, the multiclass balancing ensemble is proposed. Evaluations show that the method enhances model performance, resulting in improved classification accuracy. Good generalization properties of learned models are important and therefore a generalization study is done where models are evaluated on unseen traffic data with dissimilar traffic behavior stemming from different road configurations. This is realized by using two distinct highway traffic recordings, the publicly available NGSIM US-101 and I80 data sets. Moreover, methods for encoding structural and static features into the learning process for improved generalization are evaluated. The resulting methods show substantial improvements in classification as well as generalization performance.

Data-driven modeling and machine learning are widely used to model the behavior of dynamic systems. One application is the residual evaluation of technical systems where model predictions are compared with measurement data to create residuals for fault diagnosis applications. While recurrent neural network models have been shown capable of modeling complex non-linear dynamic systems, they are limited to fixed steps discrete-time simulation. Modeling using neural ordinary differential equations, however, make it possible to evaluate the state variables at specific times, compute gradients when training the model and use standard numerical solvers to explicitly model the underlying dynamic of the time-series data. Here, the effect of solver selection on the performance of neural ordinary differential equation residuals during training and evaluation is investigated. The paper includes a case study of a heavy-duty truck’s after-treatment system to highlight the potential of these techniques for improving fault diagnosis performance.

The ability to predict the future trajectories of traffic participants is crucial for the safe and efficient operation of autonomous vehicles. In this paper, a diffusion-based generative model for multi-agent trajectory prediction is proposed. The model is capable of capturing the complex interactions between traffic participants and the environment, accurately learning the multimodal nature of the data. The effectiveness of the approach is assessed on large-scale datasets of real-world traffic scenarios, showing that our model outperforms several well-established methods in terms of prediction accuracy. By the incorporation of differential motion constraints on the model output, we illustrate that our model is capable of generating a diverse set of realistic future trajectories. Through the use of an interaction-aware guidance signal, we further demonstrate that the model can be adapted to predict the behavior of less cooperative agents, emphasizing its practical applicability under uncertain traffic conditions.

 Given their adaptability and encouraging performance, deep-learning models are becoming standard for motion prediction in autonomous driving. However, with great flexibility comes a lack of interpretability and possible violations of physical constraints.  Accompanying these data-driven methods with differentially-constrained motion models to provide physically feasible trajectories is a promising future direction. The foundation for this work is a previously introduced graph-neural-network-based model, MTP-GO.  The neural network learns to compute the inputs to an underlying motion model to provide physically feasible trajectories. This research investigates the performance of various motion models in combination with numerical solvers for the prediction task. The study shows that simpler models, such as low-order integrator models, are preferred over more complex ones, e.g., kinematic models, to achieve accurate predictions. Further, the numerical solver can have a substantial impact on performance, advising against commonly used first-order methods like Euler forward. Instead, a second-order method like Heun's can significantly improve predictions.

Enabling resilient autonomous motion planning requires robust predictions of surrounding road users' future behavior. In response to this need and the associated challenges, we introduce our model titled MTP-GO. The model encodes the scene using temporal graph neural networks to produce the inputs to an underlying motion model. The motion model is implemented using neural ordinary differential equations where the state-transition functions are learned with the rest of the model. Multimodal probabilistic predictions are obtained by combining the concept of mixture density networks and Kalman filtering. The results illustrate the predictive capabilities of the proposed model across various data sets, outperforming several state-of-the-art methods on a number of metrics.