Keynote Lectures

From Shallow to Deep Kernel Machines
Johan Suykens, ESAT-Stadius & Leuven.AI Institute, KU Leuven, Belgium, Belgium

Human Autonomy Teaming Strategies for Physical Orchestrated Intelligence
Bogdan Epureanu, University of Michigan, United States, United States

Observability and Diagnosability of Hybrid Dynamical Systems: Theory and Applications
Maria Domenica Di Benedetto, University of L'Aquila, Italy, Italy

Abstract
Kernel methods have proven effective in many engineering domains, including system identification, control, robotics, signal processing, and decision‑making under uncertainty. Kernel-based modelling has a long tradition in machine learning, with successful applications ranging from function estimation in reproducing kernel Hilbert spaces to support vector machines and Gaussian processes. Within this landscape, least squares support vector machines (LS-SVMs) offer a versatile primal-dual characterization for classification, regression, principal and canonical component analysis, spectral clustering, recurrent modelling, approximate solutions to partial differential equations, and optimal control, among other tasks.

This talk explores a series of extensions that move from shallow to deep kernel machines and establish connections to deep architectures and generative AI. We introduce multi-level feature learning through Deep Kernel Principal Component Analysis, based on restricted kernel machines whose model representations follow from LS-SVMs via conjugate feature duality. These representations naturally relate to Restricted Boltzmann Machines and Deep Boltzmann Machines, and in the shallow case can be aligned with the evidence lower bound of variational autoencoders.

We further show how attention mechanisms in transformers correspond to asymmetric kernel functions involving distinct feature maps for queries and keys. Through a kernel singular value decomposition, this leads to a primal-dual characterization within the LS-SVM framework. The resulting Primal-Attention method provides a regularized loss formulation that enables efficient transformer training and low-rank representations.

Overall, these developments offer a unifying perspective linking kernel methods, deep architectures, and generative modelling, opening new directions for theory, algorithms, and applications.

Abstract
As autonomous vehicles transition from isolated platforms to collaborative teammates, the future of mobility will depend on Physical Orchestrated Intelligence: the coordinated fusion of human judgment, machine autonomy, and adaptive decision-making across heterogeneous teams. This presentation examines strategies for human-autonomy teaming in military and civilian operations where autonomous agents must perform dangerous tasks, respond to uncertainty, and cooperate with humans whose creativity and contextual reasoning remain essential. Although autonomous systems offer speed, endurance, and risk reduction, they are often brittle when confronted with unforeseen events. Humans, conversely, excel at improvisation but are constrained by cognitive workload, limited attention, and task saturation. Effective teaming therefore requires algorithms that can dynamically allocate responsibilities, exploit complementary capabilities, and adapt to changing mission conditions.

The presentation introduces modeling and simulation tools for training heterogeneous autonomous agents to learn task-distribution policies with humans and other autonomous systems. The technical approach extends Decentralized Partially Observable Markov Decision Processes to represent differences in sensing, communication, mobility, task capability, and risk tolerance. Reinforcement learning in synthetic environments enables agents to develop collaborative strategies, while Bayesian Online Strategy Adaptation supports resilience to open-world novelty by allowing teams to revise behaviors as unexpected conditions emerge.

A high-fidelity disaster-relief scenario demonstrates the framework in an immersive game-engine environment. A human operator interacts with autonomous teammates in real time through virtual reality, while an adaptive assistance algorithm continuously evaluates cognitive task load and provides decision support. By accounting for heterogeneous agent capabilities and human risk aversion, the system improves coordination, reduces overload, and enhances mission performance. Results indicate that trained autonomous agents can reliably collaborate with humans, dynamically redistributing tasks as conditions evolve.

This work is conducted through the University of Michigan-led Automotive Research Center, the U.S. Army Center of Excellence for modeling and simulation of ground vehicles. As the flagship academic partner of the U.S. Army Ground Vehicle Systems Center, ARC advances heterogeneous multi-vehicle teams that integrate autonomy, navigation, decision-making, intelligent power systems, and advanced structures. This research supports the next generation of adaptive commercial and defense ground systems, accelerating trustworthy autonomy at the speed of relevance and operational impact worldwide.

Abstract
Safety-critical embedded control systems, such as those in transportation and industrial plants, are becoming increasingly important as autonomy advances. These systems are naturally modeled as hybrid dynamical systems, characterized by the interaction of continuous and discrete dynamics. In their design, it is crucial to develop methods capable of dealing with such heterogeneous behaviors.

In many practically relevant scenarios, for example in automotive applications or in networked systems subject to intermittent sensing, only partial information about the system’s state is available. This makes state observability under hybrid dynamics, i.e., the possibility to reconstruct the hybrid system’s internal state from available measurements, essential for the implementation of control algorithms. Beyond this practical role, observability is also important as an intrinsic structural property of a system, underpinning the solution of problems such as fault diagnosis or detection of malicious attacks in safety- and security-critical applications.

While observability has been extensively studied in the continuous domain since the 1960s and in the discrete domain since the 1980s, its analysis for hybrid systems poses nontrivial additional challenges. In this talk, we show how hybrid features influence observability and related properties, including diagnosability and predictability, which are closely connected to safety and security. These features give rise to behaviors that cannot be directly inferred from classical results for traditional dynamical systems. To address the resulting complexity, we also discuss an approximate approach based on symbolic models, which provide finite-state abstractions of hybrid systems with guaranteed precision.

Examples drawn from applications highlight the relevance of hybrid modeling and observability analysis in safety-critical contexts.