Featured talks & speakers
As AI accelerates across transportation — from vehicle control to network management — what role remains for decades of physics-based and behavioral models? The Physics-Enhanced Residual Learning (PERL) paradigm integrates classic physics models with cutting-edge AI for transportation autonomy.
+ full abstract & bio
The talk introduces PERL's structure and fundamental theory, then showcases applications in multimodal autonomy and system-state prediction, supported by computational analysis and physical experiments on a unique automated-vehicle testbed. Ongoing work verifies "Foundation AI" outputs with physics models in agentic applications — quality-control inspection and AV evaluation — using real-world laboratory data.
Bio — Dr. Xiaopeng (Shaw) Li is the Harvey D. Spangler Professor of Civil and Environmental Engineering at UW–Madison. He founded the CATS Lab and its multi-scale CAV testbed, serves as Founding Executive Director of the Smart Highway Research Center, and directs the USDOT TRAVELS Center for rural autonomous passenger transportation. NSF CAREER awardee; PI on $40M+ in funded projects; 160+ journal papers. Interim Vice President for Education of IEEE ITSS and founding chair of the IEEE ITSS ET3 committee; ASCE Fellow; Co-Editor-in-Chief of Communications in Transportation Research. B.S. Tsinghua; M.S. & Ph.D. University of Illinois at Urbana-Champaign.
RL has emerged as a powerful method for complex control tasks — from autonomous driving to maritime navigation. Advances in value-based algorithms, including T- and K-Estimators for overestimation-bias control, deliver superior performance and convergence in Q-Learning and Bootstrapped DQN.
+ full abstract & bio
A spatial-temporal recurrent architecture strengthens autonomous ships under partial observability and maritime traffic rules; a modular DRL framework for inland-waterway surface vehicles outperforms traditional control. Work on dynamic obstacle avoidance for mobile robots and drones shows how controlled training difficulty improves generalization — reducing the Sim2Real gap across platforms.
Bio — Ostap Okhrin is Full Professor (W3) and Chair of Applied Statistics, TU Dresden. Ph.D. in Economics summa cum laude, European University Viadrina (2008). Since 2022, member of the DFG Review Board for Intelligent and Automated Traffic. His research covers reinforcement learning, data-driven decision-making, and statistical modeling for ITS, with emphasis on robustness and sim-to-real transfer. 100+ peer-reviewed publications; Associate Editor of several leading journals in statistics and computational sciences.
In-vehicle interfaces compete with driving for the driver's limited perceptual and cognitive resources. Computational driver models offer a scalable way to anticipate attention allocation and off-road glances — and to evaluate safety-relevant design choices before they reach the road.
+ full abstract & bio
Two complementary strategies are presented: data-driven models that learn multitasking and visual-sampling patterns from driving data, and first-principles models grounded in computational rationality that derive behavior from drivers' goals, bounded resources, and the driving environment — each illustrated with concrete examples for predicting attention and informing interface design.
Bio — Patrick Ebel is Assistant Professor for Computational Interaction at Hasso Plattner Institut (HPI) and University of Potsdam; previously he led a Junior Research Group at ScaDS.AI, Leipzig University. Doctorate in Computer Science summa cum laude, University of Cologne (2023), on data-driven evaluation of in-vehicle information systems. Active contributor to ACM AutomotiveUI; Early Career Best Paper Award from HFES Europe plus multiple Best Paper and Honorable Mention awards.
Collaborative driving requires automated systems to account not only for what a driver is doing but also for the range of actions they may take next. This talk explores modeling near-term driver intent as a distribution over plausible future ego trajectories, enabling downstream safety systems to account for uncertainty rather than rely on a single deterministic forecast.
+ full abstract & bio
Using automatic emergency braking as a case study, the talk shows how this can reduce unnecessary activations while preserving safety-critical behavior. This also raises a broader evaluation challenge: how do we test such systems when their actions can influence what happens next? Closed-loop world models such as GAIA-4 can support counterfactual and reactive evaluation, helping move toward safer, less intrusive, and more collaborative driving.
Bio — Flavian Pegado is a Staff Machine Learning Engineer at Wayve, where he works on the Wayve AI Driver and its end-to-end machine learning stack. He has spent the past decade in the autonomous vehicle industry, with experience spanning sensing, perception, localization, safety, and learned driving. He holds a Master's degree in Robotic Systems Development from Carnegie Mellon University's Robotics Institute and focuses on building safe, robust machine learning systems for real-world autonomous driving.
Existing AV simulators accelerate policy development yet fall short on human-AI collaboration and socially aware agent modeling. Sky-Drive bridges this with a distributed multi-terminal architecture, a multimodal human-in-the-loop framework, a continuous human-AI knowledge-exchange mechanism, and a digital-twin framework.
+ full abstract & bio
Applications span AV–human road-user interaction modeling, socially aware reinforcement learning, personalized driving, and customized scenario generation. Future directions integrate foundation models for context-aware decision support and hardware-in-the-loop testing for real-world validation — positioning Sky-Drive as foundational infrastructure for human-centered autonomous transportation research.
Bio — Zilin Huang is a Research Associate at UW–Madison and Visiting Research Associate at Purdue University. Founding member of Sky-Lab; affiliated with CCAT, the Smart Highway Research Center, and the TRAVELS Center. Ph.D. and M.S. from UW–Madison, advised by Prof. Sikai (Sky) Chen. Research spans human-centered AI, physical AI, autonomous vehicles, robotics, and ITS.
Most end-to-end VLA driving systems treat passenger commands as purely semantic, overlooking emotional states that shape comfort, trust, and acceptance. E3AD interprets free-form language, infers passenger emotion, and generates physically feasible trajectories.
+ full abstract & bio
It integrates a continuous Valence-Arousal-Dominance emotion model with a dual-pathway spatial-reasoning module combining egocentric and allocentric views, plus a consistency-oriented training strategy aligning emotional intent with grounding and action. Experiments improve visual grounding, waypoint planning, and emotion estimation — evidence for emotion-aware reasoning in safer, more human-aligned driving.
Bio — Yihong Tang is a Ph.D. candidate at McGill University and Mila, advised by Prof. Lijun Sun and Prof. Bang Liu, and a Visiting Researcher at ServiceNow Frontier AI Research (Agentic Forecasting). M.Phil. in Urban Computing, HKU; B.Eng. in Computer Science, BUPT. His work on multimodal AI agents and generative models appears in ICML, ICLR, KDD, CVPR, EMNLP, CIKM, TR-C, and IEEE T-ITS; FRQNT Doctoral Scholarship and multiple best-paper awards.
End-to-end RL learns longitudinal control directly from radar Range–Doppler representations — no hand-crafted perception features, no explicit target tracking. A PPO agent trained in a physics-inspired radar simulation regulates speed to hold a desired time-to-collision behind a stochastic lead vehicle.
+ full abstract & bio
The learned policy is evaluated on a model-scale vehicle equipped with an Infineon BGT60TR13C radar, achieving stable car-following across varying lead-vehicle speeds and operating conditions — evidence that Range–Doppler inputs alone carry the control cues needed for longitudinal vehicle control.
Bio — Paul Auerbach received his M.Sc. in applied computer science from HTW Dresden (2019) and is pursuing a Ph.D. at Barkhausen Institut, Dresden, collaborating with Prof. Ostap Okhrin's chair at TU Dresden on simulating and solving traffic scenarios with RL. His focus: realistic simulation of automotive sensors — especially radar — and transferring learned policies to physical sensors and model-scale vehicles.