Events at PATH

More Information TBA

Recording

"Data serves as a crucial foundation for research, particularly in analyzing complex systems such as transportation networks and human mobility. The primary objective of engaging with data is twofold: to gather real-world observations for model development, and to leverage these models to simulate (or generate) and better understand the dynamics of the real world. This process essentially involves learning the underlying distribution of collected samples. In recent years, Deep Generative Models, or Generative AI, have emerged as a transformative approach to learning these complex distributions directly from data. Unlike traditional statistical models, which impose parametric assumptions, DGMs train neural networks to learn the distributional characteristics and generate synthetic samples that mirror the patterns of real-world traffic and mobility data.
In this talk, I will discuss recent studies our group has been conducting to learn such distributions across multiple scales of transportation systems. We begin with TrajGAIL, a generative adversarial imitation learning framework that captures behavioral distributions of trip generation and route choices from large-scale vehicle trajectory data. Building on this foundation, TrajFlow extends the idea by applying deep generative modeling for probabilistic agent trajectory prediction. Finally, we introduce PMA-Diffusion, a physics-guided, mask-aware diffusion framework that addresses the data sparsity problem by training generative models directly on incomplete and noisy highway sensor data. Together, these studies highlight how DGMs enable scalable, physics-informed, and uncertainty-aware modeling of transportation and mobility systems."

Recording

As electric vehicles (EVs) gain a rapidly growing share of new sales, cities face a rare chance and responsibility to plan and scale a new layer of transportation infrastructure to support EV operations at scale. Specifically, public charging infrastructure (PCI) is being rolled out alongside EV adoption into a well-established road infrastructure system, creating an unusually responsive context to learn and adapt using emerging data. This talk examines how PCI is reshaping interactions among travelers, infrastructure, and communities, given its dual role at the intersection of mobility and energy systems. The presentation is organized around three questions: (1) how road users actually utilize PCI; (2) how those choices propagate through traffic patterns, reliability, and system efficiency; and (3) how PCI deployment could affect transportation system and community dynamics. By examining multiple large-scale, real-world datasets (electric taxi trajectories, personal EV charging logs, residents’ visitation patterns, and activity-location traces), I highlight what distinguishes PCI from most civil infrastructures: its strong dependence on behavioral routines, social exposure, and urban activity patterns. I will discuss where conventional planning tools often struggle to represent these dynamics, introduce modeling approaches that embed realistic preferences and habits, and discuss conditions under which expanding PCI can unintentionally reduce network efficiency or exacerbate disparities. I close with policy and research directions for building an equitable, resilient, and human-centered charging ecosystem.

Recording

Domain adaptation for object detection (DAOD) has become essential to counter performance degradation caused by distribution shifts between training and deployment domains. However, a critical factor influencing DAOD - context bias resulting from learned foreground-background (FG-BG) associations - has remained underexplored. In this work, I will present the first comprehensive empirical and causal analysis specifically targeting context bias in DAOD, investigating how FG-BG associations are encoded, their causal impact on detection performance, and their influence across domains. Using background masking, feature perturbation, class activation mapping, and a novel domain association gradient metric, we reveal that convolutional models encode FG-BG associations that undermine cross-domain generalization. Further, we trace context bias to convolutional pooling operations and propose Mask Pooling, which separates FG and BG pooling using foreground masks, significantly improving model robustness. We introduce a challenging benchmark with random backgrounds to rigorously assess DAOD robustness. Our systematic experiments highlight the importance of explicitly addressing context bias to develop more reliable and generalizable object detection systems.

Recording

As AI applications scale from research to deployment, the Big Data paradigm faces critical limitations: diminishing learning curves, unsustainable storage costs, and delayed time-to-market. This talk presents Smart-Data, a paradigm shift that combines event-driven data engineering with the Edge-Cloud continuum to extract maximum value from minimal data. We will start with the example of autonomous driving, where vehicles generate huge amount of data yet 99% proves useless for learning. We then demonstrate how intelligent triggers and distributed agents transform data pipelines. We'll explore both the vision of context-aware, federated AI systems and the concrete implementation strategies that enable scalable, sustainable autonomy, all of course from a data point of view.

More Information TBA

Recording

US policymakers increasingly focus on enabling more housing in walkable neighborhoods to reduce climate pollution, by reducing the need to drive, as well as to address the housing shortage. However, estimates of realistic pace and scale of pollution reduction have been lacking.
I will present findings from three studies overlaying geographic information on housing growth with current estimates of neighborhood residential vehicle miles traveled (VMT) modeled by Replica. First, my collaborators and I estimated that up to 70 million tonnes CO2 annually could be avoided after a decade, in an ambitious scenario where states address their housing shortages by directing housing growth to relatively low-VMT neighborhoods within each state. Second, we illustrated contrasting historical patterns of housing development around the country, highlighting five standout metros which have been building housing at a relatively fast pace since 2000 and have directed this new housing to predominantly lower VMT neighborhoods. Third, we are assessing 366 California local housing plans for their aggregate potential contribution to the State's climate targets, under the State's 6th-Cycle Regional Housing Needs Allocation and Housing Element processes. We observe modest progress in regionally allocating housing to lower-VMT cities, but potential statewide VMT reduction falls well short of the State's Scoping Plan target.
As an appendix, I will highlight Terner Center's recent study on the importance of better measuring VMT for housing policy applications."

Recording

Resilient transportation infrastructure is critical to efficient freight supply chain performance. Enabling scenario exploration, particularly under potential disruption conditions, is critical to making good decisions about resilient infrastructure investments. Yet there are limited open-source tools available to help supply chain participants and transportation planners evaluate the intersection between transportation infrastructure and passenger and freight movements. The U.S. Department of Transportation has developed two distinct tools to support these kinds of analyses, the Freight and Fuel Transportation Optimization Tool, which optimizes supply chain freight movements across a multimodal transportation network, and the Resilience and Disaster Recovery Tool Suite, which helps estimate transportation network exposure to hazards and evaluate the return on investment for resilience projects aimed at mitigating uncertain future conditions. Dr. Zhang will discuss the approaches these two tools take to enable transportation optimization and resilience analyses.

No recording available

Autonomous mobile robots are poised to transform future transportation systems by addressing a range of time sensitive and safety critical tasks. Beyond their traditional roles in logistics, search and rescue, and monitoring, robots, particularly aerial and ground systems, will be central to reshaping how goods and services move in remote areas or dense urban environments. In future smart cities, aerial and ground robots will integrate into multimodal transportation networks, enabling timely delivery of critical items such as medical supplies, consumer goods, and infrastructure components. To operate effectively in these complex settings, robots must demonstrate agile maneuvering, robust adaptation to dynamic navigation conditions, and seamless collaboration across agents and humans. In this talk, I will present ongoing research at the Agile Robotics and Perception Lab, focusing on the development of agile and collaborative machines. I will emphasize how integrating learning-based and model-driven approaches enables agility, rapid adaptation, and scalable solutions for multi-agent collaboration. These advances point toward the future of autonomous mobility, where fleets of small, resilient robots will enhance transportation, delivery, and infrastructure services in both remote areas and dense urban environments.

Recording

Mathematical programming plays a pivotal role in transportation system optimization, spanning from long-term planning to real-time operations. Traditional network optimization often overlooks the significance of endogenous travel demand. However, understanding traveler preferences and guiding their choices is crucial for enhancing system efficiency and moving towards a sustainable transportation framework. In this presentation, we explore innovative methods that integrate discrete choice modeling with optimization models. Our discussion includes various decision horizons: operations, tactics, and planning. First, we introduce a novel matching algorithm tailored for high-capacity ride-pooling systems. This algorithm, which integrates choice modeling and reinforcement learning, significantly boosts profitability by balancing between immediate acceptance probabilities and the future value of serving requests. Second, we propose a convex program that characterizes the equilibrium of strategic behavior between a transportation network company and travelers. In this model, a company can adaptively adjust pricing and routing strategies while considering travelers’ mode choices. Third, we present a convex programming formulation that can complement the traditional sequential four-step approach in travel demand forecasting. Throughout this talk, we provide numerical evidence highlighting the benefits of integrating discrete choice modeling with optimization. Additionally, we will discuss future directions and challenges in smart mobility systems.

The SMART project will establish an open data platform integrated system that provides customizable and timely accessible transit information services for individuals with special mobility needs, particularly those who are seniors and people with disabilities, bringing together data sources from various modes of providers and developing customer-oriented tools.

In this talk, I will discuss potential technical solutions for building a stochastic emulator of spatially correlated earthquake ground motions. Such an emulator could significantly enhance the credibility of seismic risk assessments for spatially distributed infrastructure systems, such as transportation and pipeline networks. The key challenge lies in the scarcity of data on spatially correlated ground motions, compounded by limited physical understanding of wave propagation process and substantial parametric uncertainty.

More Information TBA

Recording Unavailable

Review SMART Grant, RAV, and broader Freight Efforts

Recording

The past four decades witnessed the birth and growth of neurodynamic optimization, which has emerged as a potentially powerful problem-solving tool for constrained optimization due to its inherent nature of biological plausibility and parallel and distributed information processing. Despite the success, almost all existing neurodynamic approaches a few years ago worked well only for optimization problems with convex or generalized convex functions. Effective neurodynamic approaches to optimization problems with nonconvex functions and discrete variables are rarely available. In this talk, the advances in neurodynamic optimization will be presented. Specifically, In the proposed collaborative neurodynamic optimization framework, multiple neurodynamic optimization models with different initial states are employed for scattered searches. In addition, a meta-heuristic rule in swarm intelligence (such as PSO) is used to reposition neuronal searches upon their local convergence to escape local minima toward global optima. Problem formulations and experimental results will be elaborated to substantiate the viability and efficacy of several specific paradigms in this framework for supervised/semi-supervised feature selection, supervised learning, vehicle-task assignment, and model predictive control.

Recording

Knowledge of modern, state-of-the-practice traffic flow sensors provides traffic managers, researchers, and students an understanding of the operation, strengths, and limitations of current sensor technologies and enables them to make an informed decision as to which is appropriate for a particular application. Accordingly, this talk describes intrusive and non-intrusive traffic flow sensor technologies in use today, their applications and selection criteria, and typical output data. Furthermore, it provides examples of representative sensor models. The technologies discussed are mature with respect to current traffic management applications, although some may not provide the data or accuracy required for a specific application or may not perform as needed under the operational conditions anticipated at the installation site. Sensors selected for a first-time application should be field tested under conditions that will be encountered in day-to-day operation before large-scale purchases of the device are made. As alternative traffic data and information sources, such as commercial data vendors, Wi-Fi and Bluetooth sensing of smart phone locations, and connected and automated vehicle data become increasingly available, they are progressively finding their way into modern traffic management systems as a complement to conventional roadside sensors.

Recording

This talk will focus on: 1) Modeling of AV/CAV following behaviors based on ACM test data of passenger cars with different powertrains; 2) Implementing the model in network level mixed simulation for sensitivity analysis for the Co-Simulation (traffic and V2X) with NREL involving both freeway and arterials.

Recording

This presentation provides an in-depth look at our Connected Work Zone initiative—showcasing how we're using IoT, real-time data, and digital infrastructure to improve safety, awareness, and efficiency in active road work areas. We'll explore current deployments, technical components, and integration with standards like WZDx. The session will also outline our product and technology roadmap, highlighting upcoming capabilities, commercialization plans, and how our solutions scale across municipalities, contractors, and DOT partners to enhance roadway intelligence.

Recording

This talk explores how autonomous systems interact with other agents in complex shared environments, such as autonomous cars navigating alongside pedestrians and human-driven vehicles or delivery drones operating in shared aerial spaces. The first part focuses on game-theoretic planning and control, enabling robots to anticipate and influence other agents' decisions for seamless interactions. The second part examines how robots can infer the intentions and preferences of surrounding agents in multi-agent settings, extending beyond traditional inverse reinforcement learning. The talk presents mathematical theories and real-time algorithms to enhance autonomous systems' efficiency and adaptability.

Presentation Slides

Recording

Vehicle platooning systems have the potential to alleviate traffic congestion, enhance driving safety, and improve fuel efficiency. For a linear vehicle platooning with heterogeneous dynamics, an a-CACC framework is developed to achieve cooperative control while addressing delays and sensor noise disturbances. Furthermore, in scenarios involving communication interruptions, an observer-based d-aCACC controller is proposed to maintain control performance without switching to the ACC mode after degradation. The core of the linear CACC system design is based on string stability property. The feasibility of proposed CACC systems is validated through simulations and experiments. From a nonlinear perspective, an improved CACC controller is designed for a nonlinear vehicle platoon with uncertain parameters. The controller strictly follows the Lyapunov theory, ensuring uniformly global asymptotic stability (UGAS), in which state errors and parameters estimation errors asymptotically converge to zero. This approach enhances system robustness without relying on strict persistent excitation conditions.

Recording

Recording

We propose a “Universal EV Outlet” that works with an EV “carry along” charging cable—one end of the cable has a connector specific to that user’s EV, the other a plug for the Universal EV Outlet. This proposal does not interfere with, nor require change to, any existing charging stations. It does not require any new types of inlets on EVs. The components are already standardized. Eight use cases are examined to illustrate the advantages, and some limitations, of the Universal EV Outlet. The use cases illustrate how this solution: resolves the problem of multiple AC charging connectors, makes today’s “EV Ready” building codes more adaptable, lowers capital and maintenance costs, creates a solution to curbside and urban charging, increases energy efficiency, enables higher power three-phase AC charging for heavy vehicles, and facilitates use of EVs for building backup power and for vehicle-to-grid. Finally, we propose a standards-based active cable used with the Universal EV Outlet, which would allow fast and secure EV identification for curbside or other shared charging locations, usable today without modifications to current EVs.

Recording

This presentation first gives an overview of the research carried out by Prof. Yedavalli and his group on stability and robustness of dynamic systems described by linear state space models with applications in aerospace, mechanical and electrical systems using both eigenvalue based stability assessment via Transformation Compliant (TC) methods such as the Routh-Hurwitz Criterion, Cayley-Hamilton Theorem, and Lyapunov Matrix Equation methods (which are all equivalent to each other) as well as sign pattern based Qualitative Sign Stability (QLSS) approach being used by ecology researchers. Then, by juxtaposing these two extreme viewpoints, namely TC and QLSS methods, his startup firm RES proposes a new method without using eigenvalues at all. This new approach that does not depend on eigenvalues uses a new concept of stability, namely Convex Stability as opposed to Hurwitz stability (for continuous time systems) and Schur stability (for discrete time and sampled data systems) for the real state variable convergence issue of Linear Systems that include time invariant as well as time varying systems, allowing multiple equilibrium points both in continuous time and as well as discrete time domains. Thus, the Convex Stability property can handle some mild non-linear dynamic systems as well. It is shown that the new Convex Stability concept is also equivalent to the problem of Static Output Feedback (SOF) stabilization. Thus, RES CSSP Toolbox algorithms solved the long-standing problem of getting necessary and sufficient conditions for the existence of a static output feedback gain that stabilizes any LTISS system. This patented Convex Stability concept does not need the transfer function approach and proves that the celebrated Mapping theorem is not only irrelevant for convex stability (and SOF stabilization) but also is incorrect in eigenvalue-based methods. The RES CSSP Toolbox algorithms propose strict upper and lower bounds on the subharmonic frequencies (for lower bound) and super harmonic frequencies (for the upper bound) as a function of the sampling period T in power electronics circuits for stability. Similarly, it is shown that Kalman Filter state estimation errors can never possess asymptotic convergence to zero if they use current literature eigenvalue based MATLAB software routines. Because of the issuance of two patents (one an already awarded US patent # No. 11,815,862, and another India patent under review), the contents of this presentation are IP protected and can be used only with a technology licensing agreement with the firm Robust Engineering Systems, LLC or permission from it.

Recording and slides unavailable

Recording

Proposed strategy to model the required spatio-temporal charging demand from light-duty (LD) and medium- and heavy-duty (MHD) electric vehicles (EVs) using actual transportation data by mapping the demand for the required EV charging to a realistic and coordinated distribution and transmission electric grid at the predicted times of the day to study their impact on power system in a variety of load, weather and EV penetration scenarios. This work is the first study that includes the actual weather data and transportation data with realistic and coordinated distribution and transmission grid data in a large industry-scale level study. The main goal of this study is to identify possible issues and required upgrades in the electric grid, caused by an increase in EV integration. The transmission case study is a large grid with 6717 buses over Texas footprint and the distribution grid is over Houston, a city in Texas, covering over 3 million customers. The resulting overloads and voltage violations experienced in the system are discussed and required planning upgrades to avoid these issues are suggested.

Recording

Integrating Large Language Models (LLMs) into autonomous driving technology offers transformative potential across perception, decision-making, and human-vehicle interaction. In this talk, we introduce LLM4AD, a framework that leverages LLMs to enhance autonomous driving systems through natural language understanding, situational reasoning, and personalized motion control. Our system includes a novel benchmark to evaluate instruction-following capabilities and employs a Retrieval-Augmented Generation (RAG)-based memory module for continuous learning, enabling adaptive control strategies based on human feedback. Extensive field experiments using a drive-by-wire-enabled autonomous vehicle demonstrate the system’s ability to translate complex natural language instructions and environmental inputs into actionable control policies, achieving reductions in human intervention by up to 76.9% while enhancing safety, comfort, and alignment with user intent. Additionally, insights from a related study on an efficient on-board Vision-Language Model (VLM) reveal how real-world deployment of personalized motion control further bridges the gap between human preferences and autonomous vehicle behavior, setting new benchmarks in trust and reliability. Together, these advances highlight the role of LLMs in creating context-aware, responsive, and personalized autonomous driving solutions.

Recording

Recording

In this talk I focus on the dynamics and control of connected automated vehicles in mixed traffic scenarios where CAVs interact with human-driven vehicles. I will highlight the concept of connected cruise control (CCC) which enables CAVs to improve their safety and energy efficiency by responding to the motion of multiple vehicles ahead. I will show how such benefits may be retained for low penetration of connectivity using infrastructure support. The theoretical results are demonstrated using experiments with full size vehicles on test tracks and public roads.

Recording

Ensuring the safety of reinforcement learning (RL) algorithms is crucial to unlock their potential for many real-world tasks. Provably safe RL describes approaches that ensure hard safety guarantees for RL. In this talk, we will explore different conceptual approaches of provably safe reinforcement learning and demonstrate how set-based reachability analysis can be integrated with RL to obtain hard guarantees even for dynamic environments. More specifically, we will consider an action projection approach for general non-linear systems that corrects unsafe actions to the clostest safe action so that collisions with dynamic or static obstacles are avoided. Finally, we will expand the notion of safety to traffic rules and demonstrate how temporal logic specifications of traffic rules can be leveraged to ensure safety of an autonomous vessel.

Recording

Recording

The federally funded Work Zone Data Exchange (WZDx) initiative continues to advance efforts to enhance the quality of work zone data across the United States. The PATH team has been conducting a literature review on the latest WZDx Specification, the Connected Work Zone (CWZ) Standard and implementation guide, as well as example use cases from the US DOT-funded WZDx demonstration grants. Additionally, the PATH team will provide a brief overview of two related projects at PATH and outline potential future research topics.

Recording

The collective behavior of human-driven vehicles (HVs) produces the well-known stop-and-go waves potentially leading to higher fuel consumption and emissions. This letter investigates the stabilization of traffic flow via a minimum number of autonomous vehicles (AVs) subject to constraints on the control parameters aiming to reduce the number of vehicles on the road while achieving lower fuel consumption and emissions. The unconstrained scenario has been well-studied in recent studies. The main motivation to investigate the constrained scenario is that, in realistic engineering applications, lower and upper bounds exist on the control parameters. For the constrained scenario, we optimally find the minimum number of required AVs (via computing the optimal lower bound on the AV penetration rate) to stabilize traffic flow for a given number of HVs. As an immediate consequence, we conclude that for a given number of AVs, the number of HVs in the stabilized traffic flow may not be arbitrarily large in the constrained scenario unlike the unconstrained scenario studied in the literature. We systematically propose a procedure to compute the optimal lower bound on the AV penetration rate using nonlinear optimization techniques. Finally, we validate the theoretical results via numerical simulations. Numerical simulations suggest that enlarging the constraint intervals makes a smaller optimal lower bound on the AV penetration rate attainable. However, it leads to a slower transient response due to a dominant pole closer to the origin.

Recording

The potential to enable fast and reconfigurable mechanisms for increasingly prevalent networked control systems (including connected automated driving systems, ADS) underscores the critical need to develop more distributed-system designs robust to model uncertainties in time- and safety-critical scenarios. However, perception, scene understanding, reliable state estimation, and safe decision making for autonomy, is still challenging due to computational and processing constraints, uncertainties in the environment, and safety concerns. Our translational research at the Networked Optimization, Diagnosis, and Estimation (NODE) lab, University of Alberta, aims at enhancing reliability and computational efficiency of perception and motion planning for connected ADS in cooperative intelligent transportation setting. In this talk, some resilient distributed controls and state estimation methods using multi-modal remote sensing fused with onboard measurements will be presented.

Recording

The electric vehicle represents one of the most significant technological advancements of this century, driven in part by the goal of achieving carbon neutrality by 2050. However, energy storage and fuel remain major concerns for future development. While lithium-ion batteries are currently the most common energy storage solution for electric vehicles, they face challenges related to recycling and their carbon footprint. Hydrogen fuel shows promise but presents numerous hurdles, including issues with production, transportation, and safety. In contrast, metal-oxide fuel cells offer a compelling alternative for energy storage. Recently, ammonia-powered electric vehicles have gained attention due to their potential for low cost and operation under low pressure. Power electronics form the backbone of vehicle technology, encompassing critical components such as power circuits, motor drives, chargers, and power distribution systems. In this presentation, we will explore various energy storage solutions, power conversion techniques, and fuel processing technologies. Specifically, we will focus on our recent developments in ammonia-powered electric vehicles, which provide advanced features and zero emissions compared to hydrogen-based vehicles. The talk will cover the fundamental principles of ammonia power and its applications in vehicles.

Recording

The Transport Analytics research group in Linköping university, Sweden, focus on utilizing large-scale mobility data for understanding and managing transport networks. This seminar will give an overview of current research projects in the group with focus on projects related to Data-driven network assignment and Multimodal traffic management. We will also present a new innovation collaboration with the US company Xtelligent, that focus on dashboards for traffic management applications.

Recording

Recording

Scalable Safety Assessment of the State Highway System: We will review usRAP, its safety scores and how they are computed. Then, we'll explain the added value of our project, and update on its progress.