In this talk, we will present an overview of an air hockey playing robot that we built during the as a vehicle to investigate research issues in real-time control, machine learning, and computer vision. ???We will discuss the problems of supervisory hybrid control of nonlinear systems and control of systems through impacts and how they enabled a successful implementation of robotic air hockey.?

In the first part of this talk we review the current state of the art of predictive control focusing on stability, real-time feasibility and robustness. In the second part of the talk we address the problem of steering a nonlinear system along a given parametrized reference path, taking input and state constraints into account. Such problems are known as path-following problems, which typically arise in vehicle and robot control, as well as in process control. We propose an efficient predictive control scheme which guarantees convergence to the state path in the presence of state and input constraints.

Systems with uncertainty is a fundamental area of automatic control since the mathematical model of a real plant is almost never exactly known. Generally, the coefficients of such a model depend on uncertain parameters (e.g., linearly, polynomially, etc) confined in a set (e.g., hypercube, polytope, etc). Unfortunately, essential problems such as establishing whether an equilibrium point is robustly stable for all the admissible uncertainties, and determining worst-case system performances, are notoriously very difficult to solve (NP-hard problems), even in the simplest case of linear dependence of the system on the uncertainty. This talk presents our pioneering and recent results for addressing these problems via linear matrix inequalities (LMIs). Various frameworks are considered, such as continuous-time/discrete-time systems, time-invariant/time-varying uncertainties, and linear/nonlinear dependence of the system on the uncertainty. In particular, it is shown that one can build LMI conditions that are not only sufficient but also necessary for robustness investigation.

Due to increasing fuel prices, maximizing fuel-efficiency for personal and especially commercial vehicles becomes more and more relevant. Eventually, this allows for a cost-efficient usage of more complex and costly alternative powertrain setups such as hybrid electric powertrains. The control of such powertrains plays a key role in achieving the goal of maximizing fuel-efficiency for a given setup. In the case of road-going hybrid-electric vehicles, i.e. commercial long-haul trucks and passenger cars, an approach to maximize fuel-efficiency by using prediction data of the road's characteristics, e.g. road slope, is presented. In a first step an optimal control approach for offline analysis based on discrete dynamic programming is used to allow for problem specific simplifications leading to a real-time capable optimal control algorithm in a second step. The real-time capable control is subject to tests in prototype hybrid-electric vehicles.

Sensor networks are multi-agent systems consisting of distributed nodes that cooperate to meet a common objective, often in an uncertain environment. Viewed as a cyberphysical control system, a sensor network encompasses three interconnected functionalities: coverage, data source detection, and data collection. We will describe an optimization framework for controlling the location and movement of nodes so as to combine these functionalities. We also address the issue of information exchange among nodes aiming to minimize inter-node communication. We show that event-driven, rather than synchronous, communication can guarantee convergence to optimal solutions while drastically reducing the need for inter-node communication, thus also reducing energy consumption and extending the network's lifetime. Despite significant advances, these systems remain largely used for data collection from physical sources. There has been limited progress on the processing of data to drive actuation and form a complete closed-loop cyberphysical control system. We describe one such system recently developed for dynamic resource allocation in an urban environment. Termed a "smart parking" system, it dynamically assigns and reserves an optimal resource (parking space) for a user (driver) based on an objective function that combines proximity to destination and parking cost, while also ensuring that the overall parking capacity is efficiently utilized. When an optimal allocation is updated, it is guaranteed to avoid reservation conflicts and to preserve a monotonically nonincreasing cost for every user relative to current assignment. We will describe an implementation of this system at a Boston University parking facility based on driver requests entered through a smartphone app.

本講演では chance-constrained model-predictive control （CCMPC,確率制約条件つきモデル予測制御）の新たなアルゴリズムを発表する。CCMPC とは、非有界で確率的な外乱の存在下で、状態制約（state constraints）を満たせないリスクを、与えられた確率以下に抑制するモデル予測制御の技術である。本研究では、非有界外乱を受ける動的システムの、MPC による制御の問題を考える。ロバストモデル予測制御（RMPC）においては、resolvability（または recursive feasibility）はモデル予測制御器の安定性を論じる上で重要な性質であった。Resolvability とは、もし現在において制御器がfeasible な解を見つけられる場合、未来においても必ず feasibleな解を見つけることができるという性質である。しかし、外乱が非有界である場合、一般に resolvabilityを保証することは不可能である。そこで我々は probabilistic resolvabilityという新たな概念を提唱する。Probabilistic resolvability とは、もし現在において制御器が feasibleな解を見つけられる場合、向こう N ステップの未来においてある確率で feasible な解を見つけることができるという性質である。我々は新たな CCMPC アルゴリズムを提案し、それが probabilistically resolvable であることを示す。