Tutorials

Tutorial 1: Bridging Control and Learning for Safe and Scalable Autonomy: A Tour Through the Lens of Robotics

This tutorial surveys the intersection of control theory and machine learning for deploying safe, generalizable autonomous robots, structured as a progression from foundations to frontiers. It opens with a comparative benchmarking of optimal control and reinforcement learning across axes such as robustness, performance, and task generalization, then develops the theoretical underpinnings of safety-critical decision-making through frameworks including control barrier functions (CBFs), Hamilton-Jacobi reachability, and predictive safety filters, with extensions to stochastic settings. The third part addresses scalable autonomy by embedding these safety guarantees into modern learning architectures such as diffusion policies and vision-language-action (VLA) models. The tutorial concludes with a hands-on interactive session built on open-source simulation tools, allowing participants to directly explore the trade-offs between safety, performance, and generalization, with the goal of advancing reliable robot autonomy for real-world deployment.

Tutorial 2: Scientific Machine Learning for Modeling, Optimization, and Control

This tutorial introduces Scientific Machine Learning (SciML) as a unifying framework for integrating physical structure, optimization principles, and control-theoretic insights directly into learning architectures, positioned as a principled extension of classical methods rather than a replacement. Structured around four learning paradigms — learning to solve (via physics-informed neural networks), learning to optimize (for constrained optimization), learning to model (via neural differential equations), and learning to control (via differentiable predictive control) — the tutorial equips participants with both theoretical grounding and practical skills through hands-on, code-driven exercises using the open-source NeuroMANCER library and PyTorch. Emphasis is placed on building differentiable, modular pipelines that embed physical laws and constraints to achieve data-efficient learning with improved interpretability, robustness, and compatibility with safety guarantees, with the broader aim of defining a forward-looking research agenda at the intersection of learning, optimization, and dynamical systems.

Tutorial 3: Introduction to Integrated Task and Motion Planning

Task and Motion Planning (TMP) tackles a central challenge in robotics: integrating high-level task reasoning with low-level geometric and dynamic feasibility. Symbolic planners support long-horizon decision making and structured task decomposition, while motion planners ensure physical executability under kinematic, dynamic, and environmental constraints. Real-world autonomy requires principled mechanisms that tightly couple these layers, particularly under uncertainty and partial observability, while continuously incorporating sensory information and newly acquired environmental knowledge to adapt behavior online.

This tutorial introduces the fundamental concepts of TMP and provides a structured overview of modern approaches, spanning search-based integration, sampling and optimization techniques, belief-space extensions, and learning-augmented methods. We discuss hierarchical representations and decision-theoretic formulations that enable robots to reason jointly about what actions to perform and how to execute them.

The tutorial emphasizes practical design choices, scalability considerations, and emerging directions for integrating perception and learning into structured robotic planning systems, and includes hands-on exercises to explore core TMP algorithms in practice.

Tutorial 4: Data-Driven Nonlinear Modeling and Control - without Neural Networks

This tutorial introduces Spectral Submanifold (SSM) reduction, a mathematically rigorous procedure for deriving exact, low-dimensional models of high-dimensional nonlinear dynamical systems on low-dimensional, attracting invariant manifolds. Such SSMs emanate from steady states along their dominant spectral subspaces. The three-lecture program covers the theoretical foundations and broad applications of SSMs across fluid mechanics, solid mechanics, and control; their extension to infinite-dimensional delay differential equations with applications to vehicle control and traffic dynamics ; and adiabatic SSMs (aSSMs) for systems with multiple time scales, with emphasis on data-driven control of soft robots and artificial muscles. The emphasis will be on data-driven SSM- and aSSM-reduction via the open-source SSMLearn algorithm. We also show dramatic performance advantages for aSSM-reduced models when they are compared with neural-network-based models used in controlling soft robots.  The tutorial closes with a panel discussion on open challenges and future directions.

Tutorial 5: Bridging Learning-Enabled Control with Modern NN Verifier

Deep learning methods for controller synthesis have gained substantial popularity due to their high expressiveness and strong empirical performance. However, in safety-critical control applications, there is a pressing need for formal verification of such nonlinear neural network controllers. Existing verification frameworks based on semidefinite programming or polynomial optimization face severe scalability limitations for learning-based nonlinear control. This tutorial bridges learning-enabled control with α,β-CROWN, the state-of-the-art neural network verifier that achieves scalability through GPU-parallelized symbolic linear bound propagation and branch-and-bound refinement. We discuss how α,β-CROWN enables scalable parallel verification of large neural network controllers by providing a unified formulation that reformulates diverse control verification problems such as Lyapunov stability, barrier-function safety, contraction-metric analysis, and dissipativity-based robustness as verification conditions directly solvable by α,β-CROWN. We will also discuss emerging training frameworks that jointly synthesize controllers and verifier-friendly certificates. The tutorial will include both theoretical discussions and hands-on experience with interactive programming tutorials. The aim of this tutorial is to systematically introduce α,β-CROWN to the community as a fresh perspective to solve complex challenges in verifying nonlinear and learned controllers. Coding examples are available at https://github.com/Verified-Intelligence/abCROWN_Control_Tutorial  

Tutorial 6: Qualitative and Quantitative Verification of Neural Networks and Learning-enabled Cyber-Physical Systems with StarV

Ensuring that artificial intelligence (AI) systems are safe, secure, and trustworthy is critical, particularly in safety-critical applications such as learning-enabled autonomy. Achiev-

ing these objectives for reliable next-generation AI requires formal verification, particularly for machine learning (ML) components such as neural networks. This tutorial introduces StarV, a novel software verification tool that enables both qualitative and quantitative verification for neural networks and learning-enabled autonomous Cyber-Physical Systems. This interactive tutorial begins with an overview of neural network verification research, followed by hands-on demonstrations using StarV and real-world examples.