The Schedule

 The increasing complexity of high-dimensional interconnected systems in optimization and control necessitates scalable methods that combine strong theoretical guarantees with practical performance and efficient implementability. Typical application examples for these methods range from distributed optimization in large-scale machine learning problems to cooperative control tasks such as coordinated formation control of autonomous vehicles. Traditionally, most distributed solution approaches are duality-based and therefore rely on strict assumptions such as strong duality or convexity and entail a substantial communication effort. To address these challenges, this talk presents a method termed sensitivity-based distributed programming (SBDP) for the distributed solution of nonlinear optimization and optimal control problems. The method is inherently applicable to cooperative, distributed model predictive control which is a particularly challenging application example of distributed optimization. SBDP fundamentally differs from duality-based schemes such as dual decomposition or the alternating direction method of multipliers (ADMM). The scheme exploits local sensitivity information of neighboring subsystems to ensure coordination, while maintaining scalability through a primal decomposition framework. Its design avoids the overhead of reformulating the original problem as a generalized consensus problem and instead iterates directly on the coupled variables. As a result, the approach is applicable to a larger problem class, yields subproblems of smaller dimension, reduces communication requirements, and improves scalability compared to dualitybased approaches. The talk focuses on the development, algorithmic analysis, and evaluation of this approach. It demonstrates that the method is competitive with state-of-the-art algorithms, such as ADMM, both from a theoretical and a practical perspective.

In this talk I will present a distributed learning and optimization framework in which agents cooperate to perform global tasks in uncertain operating conditions while learning knowledge on the environment. For this scenario, I will present a distributed feedback framework in which optimization problems are concurrently solved with data-driven learning updates and applied to the real system. The proposed framework is hierarchical in the sense that it handles simultaneously algorithmic and (possibly) physical dynamics thanks to a proper feedback interconnection of the different modules. Moreover, I will present a multi-resolution approach in which agents interact locally according to  distributed updates to learn a macroscopic model of the multi-agent system and achieve a desired emergent global behavior. Finally, I will show applications of the above feedback optimization and learning framework to cooperative multi-robot systems carried over thanks to novel ROS 2 (Robotic Operating System 2) toolboxes for distributed robotics.  

In this talk, we consider state estimation of nonlinear multi-agent systems with potentially dynamic network architecture. In the first part, we derive practical tools to verify detectability of the overall network by analyzing the small-scale subsystems and their interactions with its direct neighbors, utilizing a small-gain condition. In the second part of the talk, we solve the estimation problem using a distributed moving horizon estimation (MHE) approach. To ensure low communication and computation requirements, we design a small-scale MHE scheme to estimate the agents’ states, relying only on the locally available sensor data and data provided by the directly neighboring agents. Under a small-gain condition, we derive theoretical guarantees for the state estimate of the overall network, ensuring stability and robustness of the overall estimation error under practical and verifiable conditions.

Distributed optimization algorithms at the core of many applications in control such distributed MPC or federated learning. However, the performance of numerical optimization algorithms depends heavily on the tuning parameters and on the chosen initialization. In this talk, we will first give an overview on recent progress on distributed non-convex and convex optimization for control. Then we will focus on how machine learning techniques can be leveraged to speed up distributed optimization algorithms. We will consider the problem of warm starting dual decomposition for MPC problems. We show how tailored kernel regression can be leveraged to improve the performance of dual decomposition in certifiable manner. We draw upon several examples to support our findings. 

This talk will present recent works at the intersection of distributed model predictive control and multi-agent reinforcement learning, with a focus on learning efficiency and computational tractability. We begin by introducing a novel framework that uses distributed MPC as a function approximator for value functions in multi-agent reinforcement learning for constrained linear systems. By carefully structuring the MPC parameterization, we show how Q-learning updates can be performed distributively to preserve a fully distributed implementation.

Building on this foundation, we then move beyond first-order learning methods and present a second-order extension of MPC-based distributed Q-learning. Leveraging locally available second-order information and neighbor-to-neighbor communication, this approach enables significantly faster convergence.

Finally, we broaden the scope to distributed control of piecewise affine systems. We introduce a distributed MPC strategy that avoids mixed-integer optimization by solving only convex problems, using an ADMM-based method to handle the nonconvexities induced by PWA dynamics.

Multi-agent robotic systems are often required to achieve collective objectives such as coverage and coordination while accounting for nonlinear dynamics, safety constraints, and limited communication. This talk reviews how model predictive control (MPC) provides a natural framework for addressing these challenges, with applications ranging from coverage control and communication-preserving coordination to scenarios with position-dependent communication and time-varying interaction graphs.
We begin by discussing distributed MPC formulations based on distributed optimization, in which agents cooperatively solve the optimal control problem using only neighbor-to-neighbor communication. We highlight both the strengths and the limitations of these approaches.
We then turn to robotic settings in which coupling arises through position-based constraints and partition-based cost functions. In these cases, the problem can be naturally formulated as distributed non-iterative control using contracts: neighboring agents exchange predicted trajectories or set-based summaries of their future motion, allowing coupled constraints to be enforced locally. This perspective enables scalable, non-iterative coordination, naturally accommodates time-varying position-dependent interaction graphs as agents join or leave, and helps preserve communication within the swarm during closed-loop execution.

Cooperative collision avoidance between autonomous robots, or `agents,' in swarm operations remains an open challenge. Assuming a decentralized architecture, each agent is responsible for making its own decisions and choosing its control actions. For avoiding collisions, most existing approaches rely on a (wireless) communication network between (some of) the agents. In reality, however, communication is brittle. It may be affected by latency, further delays and packet losses, and transmission faults. Moreover, it is subject to adversarial attacks, such as jamming or spoofing. 

In this talk, we explore possible approaches for collision avoidance in robot swarms that do not rely on communication. The idea is that all agents base their control algorithm on a few universal rules that are agreed offline. A suitable analogy is road traffic, where simple traffic rules allow vehicles and other traffic participants to pass each other safely, even very closely and at high speeds.

For instance, the universal rules may include the definition of a specific contingency trajectory for each robot. The collision avoidance constraints may be constructed using perpendicular bisecting planes. This setup of Contingency Model-based Control (CMC) permits a full guarantee of recursive feasibility and collision avoidance between all swarm members in closed-loop operation. CMC naturally satisfies the plug-and-play paradigm, i.e., new robots may enter the swarm dynamically. The effectiveness of the CMC regime is demonstrated in numerical examples, showing that the collision avoidance guarantee is intact and the robot swarm operates smoothly in a constrained environment.