Upcoming Rising Star Talks

Abstract: Autonomous robots need a metric state to reliably interact with the physical world, yet current state estimation methods can become unreliable under poor initialization, limited excitation, and sensing degeneracy. This talk asks a simple question: when can a robot certify its estimated state? Using GNSS/INS integration as a concrete testbed, I discuss three challenges. First, what are the practical limits of widely used estimators such as EKF and factor graph optimization? Second, can nonlinear state estimates be accompanied by certificates of global optimality? Third, when the sensing problem becomes degenerate, what can still be estimated and certified? I will present our recent work on estimator comparison, certifiable optimization, and degeneracy-aware estimation, and show how these ideas can provide stronger guarantees for metric state estimation. More broadly, this work aims toward certifiable metric state estimation for autonomous robots, connecting spatial and temporal estimation with optimization and certification.

Bio: Baoshan Song is a Ph.D. candidate in the Intelligent Positioning and Navigation Lab (IPNL) at The Hong Kong Polytechnic University. His research focuses on state estimation and optimization for autonomous robots, with an emphasis on certifiable metric state estimation. His work combines GNSS/INS integration, nonlinear optimization, and semidefinite programming to study how robots can obtain reliable metric states under challenging sensing conditions. He has worked on robot localization, sensor calibration, factor graph optimization, and globally certifiable estimation. His broader research goal is to develop estimation methods that not only produce accurate states, but also provide computational evidence about their optimality and observability.

Abstract: As autonomous agents are rapidly being integrated into everyday process - assistive robotics, autonomous driving, aviation, industrial robotics, and co-working agents - there is a critical need to understand  how should agency be allocated and adapted between humans and AI across contexts and timescales?

Existing approaches remain fragmented, lacking a principled approach in developing a theoretical framework for shared autonomy. Robotic systems rely on control blending heuristics; AI assistants employ implicit delegation; emerging agentic systems lack principled authority structures. These methods fail to account for the co-evolving dynamics defining human-AI collaboration.

This work introduces a shared authority framework, termed Flip-Team, a game-theoretic approach to authority switching, to learning-based settings where system dynamics or cost structures are unknown. The framework under a cooperative setting result as an identical-interest dynamic game, yielding optimal switching policies through closed-form value function recursions for linear-quadratic systems. However, real-world human-AI interactions often involve complex, nonlinear dynamics that preclude analytical solutions. We address this limitation by learning the value function from interaction data, enabling optimal switching in systems where the underlying dynamics must be estimated online.

The key insight is that authority switching can be cast as comparing learned value functions against switching cost thresholds. When the expected cost-to-go under AI control exceeds that under human control by more than the switching penalty, authority should transfer. By parameterizing the value functions as neural networks and training on observed state-action-cost trajectories, we obtain switching policies that generalize across states without requiring explicit system identification. Preliminary results on simulated tasks demonstrate that learned switching policies outperform fixed-authority baselines. This work provides a unified framework for human-AI collaboration where authority is dynamically allocated based on context, competence, and cost.

Bio: Sandeep Banik is a Postdoctoral Research Associate at the University of Illinois Urbana-Champaign, working with Prof. Naira Hovakimyan. His research develops game-theoretic and control-theoretic foundations for human-AI collaboration and shared autonomy, including principled authority switching between humans and AI agents, decision-making under distribution shift, and multi-agent coordination for human-robot teaming. He received his Ph.D. in Electrical and Computer Engineering from Michigan State University in 2023, where his dissertation focused on adversarial modeling in game-theoretic frameworks for securing cyber-physical systems. He also holds an M.Sc. in Mechanical Engineering from KU Leuven and a B.Eng. from Visvesvaraya Technological University. His work has appeared in venues such as IEEE Transactions on Automatic Control, IEEE Transactions on Robotics, CDC, IROS, and RO-MAN, and he received the IEEE TCSP Best Student Paper Award at CDC 2023.

Abstract: Extended reality (XR) applications are moving beyond entertainment into medicine, education, manufacturing, and defense. The head-mounted displays that run them are sensor-rich systems, equipped with IMUs, depth sensors, and eye and hand tracking that continuously capture the user's body movements and physical surroundings, while immersive interfaces mediate the user's entire perception and interaction. This human-in-the-loop coupling creates a new attack surface, where compromising the interface can directly manipulate user behavior and undermine their safety and privacy. This talk presents my work on characterizing and defending against these threats. First, I introduce user interface (UI) attacks on the immersive web (WebXR), showing how adversarial content can occlude and manipulate 3D interfaces to drive users into unintended, security-critical actions while they remain largely unaware. Second, as existing 2D UI defenses do not directly translate to XR, I present EXRPOSE, a runtime defense that formalizes security policies in temporal logic and detects UI-based attacks in WebXR as they occur. Together, these contributions provide a path toward secure user interaction in XR. I will conclude with open problems and my future directions in securing sensor-rich, human-in-the-loop immersive systems.

Bio: Chandrika Mukherjee is a Ph.D. student in Computer Science at Purdue University, advised by Prof. Z. Berkay Celik in the PurSec Lab. Her research focuses on the security and privacy of emerging mobile systems, particularly extended reality (XR) headsets. She combines mixed-method user studies, system design, formal methods, and machine learning to understand threats to the users and developers of these systems and to build practical defenses against them. Her work has appeared at USENIX Security, ACM CCS, NDSS, and PETS, has received an Honorable Mention at USENIX Security 2025 and the Andreas Pfitzmann Best Student Paper Award at PETS 2026, and has been featured in WIRED. She has completed three internships at Meta, working on privacy and machine learning infrastructure. She received her M.S. from Purdue University and her B.Tech. in Computer Science and Engineering from NIT Durgapur, India.

Abstract: Accurate and reliable positioning in dense urban environments remains a major challenge for Global Navigation Satellite System (GNSS)-based Real-Time Kinematics (RTK) due to multipath and Non-Line-of-Sight reception. Cooperative positioning has emerged as a promising approach to address these challenges by allowing nearby agents to exploit shared and relative information. Agents operating under poor signal conditions can benefit from neighbouring agents with more favourable observation geometries.

This talk introduces the concept of multi-agent cooperative RTK (C-RTK) and provides insights into its potential benefits based on simulation studies using realistic GNSS measurements and vehicle geometries. The mechanisms behind performance improvements are discussed, with a particular focus on the role of float ambiguities and integer ambiguity resolution. These findings are confirmed using real-world GNSS observations collected in urban environments in Germany. Beyond demonstrating the potential of C-RTK, the talk explores the limits of cooperation and discusses the conditions under which additional relative information result in improved positioning performance — and when it may not provide the expected benefit.

Bio: Anat Schaper received her Bachelor and Master of Science degrees in Geodesy and Geoinformatics from Leibniz University Hannover (LUH) in 2020 and 2021, respectively. Since 2022, she has been pursuing a PhD in the Positioning and Navigation Research Group at the Institut für Erdmessung (Institute of Geodesy) at LUH. Her research interests include GNSS signal modelling, cooperative navigation in urban areas, multi-sensor fusion, and Kalman filtering for relative positioning. She is recipient of the Leo-Brandt-Award "DGON Master of Navigation 2022", and Outstanding Presentation Award at European Navigation Conference (ENC) 2025. She has been part of the core team of the Young Navigators, a network for students and young professionals in the field of Positioning, Navigation and Timing, since 2026.