Tutorials

No.	Tutorial Title	Organizer
1	Image-Based 3D Model Reconstruction: From Multi-View Geometry to Neural Rendering	Prof. Soohwan Song (Dongguk University)
2	Fun and Foundations of Swarm Systems and Control	Prof. Masaki Ogura (Hiroshima University)
3	Dynamically Embedded Model Predictive Control: Real-Time Operation, Constraint Satisfaction, and Stability	Prof. Hyungbo Shim (Seoul National University)
4	Learning-Based Control for Automotive Systems: From Semi-Parametric Dynamics Modeling to Online Policy Improvement	Prof. Heejin Ahn (KAIST)
5	From AI & Robotics Research to Railway Deployment	Dr. Calire Nicodeme (ALSTOM)

Image-Based 3D Model Reconstruction: From Multi-View Geometry to Neural Rendering

Time	Program	Speaker
09:00 – 09:50	Multi-View Stereo Theory and Practice: COLMAP / DUST3R / VGGT	Prof. Soohwan Song (Dongguk University)
09:55 – 10:05	Break
10:05 – 11:00	3D Gaussian Splatting Theory and Practice, and Case Studies in Diffusion-Based 3D Generation

This tutorial provides a comprehensive introduction to state-of-the-art techniques for photorealistic 3D model reconstruction from multi-view images. It begins with the core principles of Multi-View Stereo (MVS), a foundational technique in image-based 3D reconstruction, and then guides participants through hands-on exercises using widely adopted open-source tools such as COLMAP, DUST3R, and VGGT. The tutorial further introduces 3D Gaussian Splatting (3DGS), a rapidly advancing neural rendering technique, with practical sessions focused on building high-quality 3D rendering models. Finally, it discusses recent advances that combine generative AI, particularly diffusion models, with 3D modeling for the creation of novel 3D assets.

Fun and Foundations of Swarm Systems and Control

Time	Program	Speaker
09:00 – 09:10	Opening & Introduction	Organizer
09:10 – 09:40	Understanding and Control of Fish Schooling	Prof. Naoki Wakamiya (The University of Osaka)
09:40 – 10:10	Shepherding Dynamics: A Testbed for Nonlinear and Information-Limited Multi-Agent Control	Prof. Masaki Ogura (Hiroshima University)
10:10 – 10:20	Break
10:20 – 10:50	Swarm Control of Biohybrid Robots: Challenges and Opportunities	Prof. Yang Bai (Hiroshima University)
10:50 – 11:20	From Early Warning to Re-stabilization: Dynamical Network Markers for Critical Transition Control	Dr. Hampei Sasahara (The University of Tokyo)

Understanding and control of fish schooling

This tutorial introduces recent advances in understanding and controlling a biological collective behavior known as fish schooling. Many fish species spontaneously self-organize themselves to swim in coordination like a single entity, providing benefits such as energy efficiency, improved foraging, and enhanced predator avoidance. Their collective behavior emerges without centralized control: individual fish rely on simple behavioral rules based on local information and respond to their surrounding environment including neighbor fishes. This distributed nature of fish schooling offers valuable insights for the design of scalable, adaptive, and robust multi-agent systems. In particular, control through environmental modulation, rather than precise manipulation of individual agents, enables emergence of desired global outcomes from local interactions and environmental responses. The talk will present examples of mathematical modeling of schooling behavior of living fishes and their validations.

Shepherding Dynamics: A Testbed for Nonlinear and Information-Limited Multi-Agent Control

The shepherding problem focuses on guiding a collective of swarm agents using one or more external agents (“sheepdogs”). Since its introduction, shepherding has become a foundational model in swarm control due to its simplicity, universality, and applicability. Beyond its biological origin, shepherding dynamics naturally serve as a testbed for nonlinear and information-limited multi-agent control, where classical linear or convex tools cannot be directly applied. This talk reviews the early development of shepherding and highlights recent advances in handling heterogeneity within swarms, coordinating multiple shepherds under limited or no communication, and performing shepherding using restricted sensing such as bearing-only measurements. These developments provide insights that may inform the design of robust and scalable control strategies for multi-agent systems.

Swarm Control of Biohybrid Robots: Challenges and Opportunities

Biohybrid robots, or cyborg systems, combine living organisms with artificial sensing, computation, and stimulation devices. They can be programmed and controlled like robots, while retaining the mobility, energy efficiency, adaptability, and robustness of biological organisms. These properties make cyborg systems promising for operation in complex and unstructured environments. While most existing studies focus on single-agent control, many real-world applications require multiple cyborg agents to work together. Swarm control of biohybrid robots involves two fundamental challenges: biological agents may respond differently to the same stimulation, and electrical stimulation should be minimized to preserve their natural autonomy and capabilities. This talk presents recent efforts toward a general swarm control framework for cyborg systems. The central idea is to coordinate biological agents without treating them as fully controllable machines. Instead, biohybrid swarm control should make use of the natural behavior of living agents and guide the swarm only when necessary. Experimental results on cyborg robot swarms demonstrate that this framework can support diverse coordination tasks, including navigation, transport, and formation control. Overall, this talk highlights a shift from enforcing biological agents to coordinating with them, opening new opportunities for scalable, adaptive, and resilient swarm intelligence in biohybrid robotic systems.

From Early Warning to Re-stabilization: Dynamical Network Markers for Critical Transition Control

Critical transitions are abrupt changes in system behavior that arise in a wide range of complex networked systems, including disease onset in biological networks and cascading failures in power systems. For high-dimensional systems, however, model-based stabilization is often hindered by the limited availability of data, such as high-dimensional low-sample-size (HDLSS) datasets and snapshot observations, making conventional system identification challenging. This tutorial presents a dynamical network marker (DNM)-based framework that links early warning signals of critical transitions to control-oriented intervention. By exploiting dynamical information embedded in DNM structures, the framework enables the construction of dynamical models around critical regimes and the design of re-stabilization strategies via pole shifting. The tutorial introduces the underlying theory, methodological foundations, and illustrative examples, demonstrating how DNM-based approaches can support re-stabilization of complex networked systems before critical transitions occur.

Dynamically Embedded Model Predictive Control: Real-Time Operation, Constraint Satisfaction, and Stability

Time	Program	Speaker
09:00 – 09:10	Opening & Tutorial Overview	Prof. Hyungbo Shim (Seoul National University)
09:10 – 09:30	Motivation & Problem Formulation	Prof. Jin Gyu Lee (Seoul National University)
09:30 – 10:00	Constrained Optimization from a Control Perspective	Dr. Yong Joo Do (Seoul National University)
10:00 – 10:10	Break
10:10 – 10:40	Soft-Constrained MPC: Feasibility and Stability	Dr. Hyeonyeong Jang (Seoul National University)
10:40 – 11:10	Dynamically Embedded MPC for Real-Time Operation	Prof. Hyungbo Shim (Seoul National University)
11:10 – 11:20	Break
11:20 – 11:50	Constraint Satisfaction under Inexact Optimization	Dr. Hyungjo Byun (Seoul National University)
11:50 – 12:10	Simulation in Vehicle Control and Practical Considerations	Dr. Jinsung Kim (Hyundai Motor Company)

This talk motivates the need for reliable real-time implementation of constrained Model Predictive Control (MPC). Conventional MPC assumes its optimal control problem stays feasible and is solved to sufficient accuracy within every sampling interval. In fast or complex systems, however, limited computation can produce inexact control inputs, while abrupt reference changes, disturbances, and model mismatch can cause infeasibility and interrupt the control update. Based on these challenges, the tutorial focuses on three objectives: continuous generation of control inputs in real time, systematic management of feasibility and constraint satisfaction, and closed-loop stability guarantees. It concludes by showing how constrained optimization from a feedback-control perspective, soft constraints, dynamically embedded MPC, reference governance, and constraint tightening address these objectives.

This talk introduces Controlled Multipliers Optimization (CMO), a framework that interprets constrained optimization from a control perspective. In CMO, Lagrange multipliers are viewed not only as dual variables but also as control inputs acting on continuous-time optimization dynamics, while constraint residuals are treated as outputs to be regulated. Starting from equality-constrained optimization, the talk explains how proportional-integral feedback and feedback linearization can be used to construct multiplier dynamics that drive constraint residuals to zero and guide the optimization variables toward stationary points. It also discusses the relationship between CMO and primal-dual gradient dynamics, including the effect of proportional gain on convergence behavior. This perspective provides a conceptual bridge between optimization algorithms and feedback control, and establishes a foundation for the subsequent treatment of an optimizer as a dynamical system interconnected with the plant.

This talk introduces soft-constrained MPC for reference tracking as a systematic approach to mitigating infeasibility while retaining stability guarantees. State and terminal constraints are relaxed using slack variables, while hard input constraints are preserved and an artificial steady state is incorporated into the tracking formulation. These modifications enlarge the feasible region and permit controlled constraint violations when the corresponding hard-constrained problem is infeasible. The talk explains the roles of the slack penalties, terminal cost, terminal set, and artificial steady state, and presents the conditions used to establish asymptotic stability and input-to-state stability under additive disturbances.

This talk presents dynamically embedded MPC, in which an evolving estimate of the solution to an optimal control problem is embedded in the internal state of a continuous-time dynamic controller running in parallel with the plant. Starting from the MPC formulation, the Lagrangian and Karush-Kuhn-Tucker residuals are used to construct projected primal-dual optimization dynamics, allowing continuous control-input generation without waiting for exact solver convergence at each sampling instant. The coupled plant-optimizer system is then analyzed using input-to-state stability and a small-gain argument, which clarifies why the optimization dynamics must be sufficiently fast relative to the plant. The talk also provides a step-by-step procedure for constructing the dynamic controller from the optimal control problem and discusses design choices including initialization and optimizer gain.

This talk examines how constraint satisfaction can be maintained when real-time MPC operates with optimization error or experiences abrupt reference changes. An Explicit Reference Governor (ERG) generates an auxiliary reference from the optimization residual and available safety margin, thereby moderating command changes when the optimizer cannot track its solution rapidly. Dynamic Constraint Tightening (DCT), in contrast, reserves constraint margin to absorb the input error produced by a finite number of optimization iterations and updates the tightening level as the plant state changes. Finally, the talk explains how these updates balance performance, conservatism, recursive feasibility, and stability.

This talk illustrates the tutorial concepts through a vehicle-control simulation. A modular path- and yaw-rate-tracking environment combines linear time-invariant, linear time-varying, and nonlinear prediction models with conventional optimization solvers and a primal-dual-gradient-flow-based dynamic optimizer. Using common tracking costs, constraints, prediction horizons, and evaluation rules, the simulations compare optimization algorithms and prediction models. Through these comparisons, the talk shows how the choice of optimizer and prediction model affects control performance. It also discusses practical choices for the vehicle model, sampling and prediction horizons, constraints, computational budget, and performance metrics, as well as future use of an Explicit Reference Governor (ERG) and Dynamic Constraint Tightening (DCT).

Learning-Based Control for Automotive Systems: From Semi-Parametric Dynamics Modeling to Online Policy Improvement

Time	Program	Speaker
13:00 – 13:10	Opening & Introduction	Organizer
13:10 – 14:25	Learning-Based Modeling of Vehicle Dynamics: Parametric, Non-parametric, and Semi-parametric Approaches	Prof. Cheolhyeon Kwon (UNIST)
14:25 – 14:35	Break
14:35 – 15:50	Model Predictive Control as Online Policy Improvement: A Reinforcement Learning Perspective	Prof. Kyunghwan Choi (KAIST)

Learning-Based Modeling of Vehicle Dynamics:
Parametric, Non-parametric, and Semi-parametric Approaches

This tutorial covers different learning spectrums of vehicle dynamics modeling for trajectory-tracking control. We begin with parametric models, ranging from kinematic and dynamic bicycle models to higher-fidelity vehicle dynamics models, along with the offline and online techniques for identifying their parameters from data. We then turn to non-parametric models, principally Gaussian process regression, and their training pipelines and scalability limits. This motivates semi-parametric (grey-box) models, which augment a physics-based nominal model with a learned residual so that structure and data compensate for each other’s weaknesses. We review state-of-the-art methods trained both offline and through online iterative learning, demonstrated within the autonomous racing envelope.

Model Predictive Control as Online Policy Improvement:
A Reinforcement Learning Perspective

This tutorial presents model predictive control (MPC) as a practical framework for online policy improvement in learning-based control. We first revisit the relationship between optimal control, reinforcement learning, and policy improvement, and then show how an offline-trained policy can be refined online through finite-horizon MPC using model predictions, objective functions, and constraints. We further discuss how MPC-generated improved actions can provide training targets for updating an actor network, connecting online optimization with policy learning. The tutorial will combine theoretical insights with vehicle control examples to provide an intuitive understanding of MPC-based online policy improvement.

From AI & Robotics Research to Railway Deployment

Time	Program	Speaker
14:00 – 14:15	Opening & Introduction	Organizer
14:15 – 14:35	Railway Robotics Use Cases & System Architecture	Claire NICODEME
14:35 – 15:05	Perception in Railway Robotics	Claire NICODEME & Daisy CHAPMAN CHAMBERLAIN
15:05 – 15:15	Break
15:15 – 15:40	Drones for Railway Inspection and Monitoring	Matthieu LEVEQUE
15:40 – 16:05	Telecommunications & Connectivity Constraints	Marc FOURNIER
16:05 – 16:15	Break
16:15 – 16:30	Ethics, Security & Human Factors in Deployment	Claire NICODEME & Marc FOURNIER & Daisy CHAPMAN CHAMBERLAIN
16:30 – 16:40	From Research to Deployment – Lessons Learned	Claire NICODEME
16:40 – 17:00	Q&A	All

This tutorial bridges the gap between academic AI and robotics, and their application to industrial railway systems. It explains which robotic technologies are actually deployed in industry and how industrial constraints reshape algorithmic choices. A strong emphasis is placed on perception, highlighting its central role in task automation within harsh, repetitive, and evolving industrial environments, where robustness and human supervision are prioritized over raw performance.