July 8-11, 2014, Lisbon, Portugal
Scientific chair: João Xavier
The program contains both overview and focused lectures, and covers hot research topics in distributed multi-agent systems from signal processing, control, and optimization:
July 8, 9h00-13h00
Data Science: Rethinking DSP
July 8, 15h00-18h00
Cooperative Navigation and Control of Networked Autonomous Vehicles
July 9, 9h00-12h00
Sequential Decision-Making Under Uncertainty: an Introduction (MDPs, POMDPs, Reinforcement Learning)
July 9, 14h00-16h00
Multi-agent Decision Making Under Uncertainty: Theory and Practical Applications
July 10, 8h30-10h30
Cooperative Robot Localization and Target Tracking Based on Least Squares Minimization
July 10, 11h00-13h00
Multiple Vehicle Motion Planning: an Infinite Dimension Newton Optimization Method
July 10, 15h00-17h00
Distributed Inference in Wireless Sensor Networks
July 11, 8h30-10h30
Distributed Optimization in Multi-agent Systems
July 11, 11h00-13h00
Communication-Efficient Algorithms for Distributed Optimization
July 11, 15h00-17h00
Cooperative Estimation Techniques for Formations of Autonomous Vehicles
Cooperative Estimation Techniques for Formations of Autonomous Vehicles
July 11, 15h00-17h00
Abstract: This talk addresses several problems in the area of cooperative estimation in multi-vehicle systems. We consider the case in which only a subset of the agents have access to inertial position measurements, and the rest must rely only on local measurements and limited communication with neighboring agents to estimate their own position and velocity. We present a decentralized solution based on relative position measurements between the vehicles which features global stability and H2 performance guarantees for an arbitrary fixed measurement graph, and extend this framework to the case where the measurement topology changes over time. We also present a range-based cooperative navigation solution for an intervention autonomous underwater vehicle (I-AUV) and autonomous surface craft (ASC) tandem, developed for the EU project TRIDENT.
Bio: Daniel Viegas is a researcher at the Dynamical Systems and Ocean Robotics laboratory (DSOR) at the Institute for Systems and Robotics (ISR) in Lisbon, Portugal. He concluded his MSc in Electrical Engineering at Instituto Superior Técnico (IST) in 2010, and has been working towards a PhD degree at DSOR since 2011. His research interests include decentralized state estimation, multi-vehicle systems, observer and estimation theory for nonlinear systems, and Kalman filtering techniques.
Communication-Efficient Algorithms for Distributed Optimization
July 11, 11h00-13h00
Abstract: We design distributed algorithms for solving optimization problems. The scenario is a network where each node has exclusive access to a cost function, and the goal is to minimize the sum of all functions. We assume each function depends on arbitrary components of the optimization variable. While this makes the problem very challenging, it also enables modeling a wide variety of problems in signal processing, machine learning, and control theory. Our algorithms are distributed in the sense that no central node is used, and communications occur exclusively between neighboring nodes. In spite of being very generic, our algorithms are communication-efficient and outperform prior distributed algorithms, even ones that were designed for specific applications. We illustrate their efficiency with experiments for average consensus, distributed compressed sensing, distributed support vector machines, and distributed model predictive control.
Bio: João Mota is a Research Associate at the University College London, UK. He obtained his PhD from Carnegie Mellon University, PA, and from Instituto Superior Técnico, Technical University of Lisbon, Portugal, in 2013, and his MSc from Instituto Superior Técnico, Technical University of Lisbon, Portugal, in 2008. While working towards his degrees, he was affiliated with Institute for Systems and Robotics, Lisbon. His research interests include compressed sensing, distributed and parallel methods for information processing, sensor networks, and image processing.
Distributed Optimization in Multi-agent Systems
July 11, 8h30-10h30
Abstract: Recently, there has been a strong interest and progress in distributed optimization for multi-agent systems, motivated by applications in sensor and multi-robot networks, as well as in distributed learning and big data analytics. A typical studied setup assumes a generic, connected network of agents, whereby each agent i holds a private, convex cost function, known only by i, and the goal is to minimize the overall sum of the agents' costs subject to a (vector) variable of common interest. For the above problem, several distributed algorithms have been recently proposed, including (sub)gradient, augmented Lagrangian, and alternating direction methods. In this lecture, we present these state-of-the art methods, explain how they are constructed, how we can analyze their convergence and convergence rates, and present techniques on how the methods can be accelerated. We illustrate the algorithms on several applications in sensor networks and distributed learning.
Bio: Dušan Jakovetić obtained a dipl. ing. degree from the School of Electrical Engineering, University of Belgrade, Serbia, in August 2007, and the PhD degree in electrical and computer engineering from Carnegie Mellon University, Pittsburgh, PA, and Instituto de Sistemas e Robótica (ISR), Instituto Superior Técnico (IST), Lisbon, Portugal, in May 2013. From June to September 2013, he was a postdoctoral researcher at IST. Since October 2013, he is a research fellow at the BioSense Center, University of Novi Sad, Serbia. His research interests include optimization and signal processing for distributed systems and sensor networks.
Distributed Inference in Wireless Sensor Networks
July 10, 15h00-17h00
Abstract: Distributed inference can be broadly described as the set of algorithms and methods in which a number of individual agents cooperate and coordinate to achieve a certain detection, estimation or tracking task. Any application involving wireless sensor networks has, at its core, a certain distributed inference algorithm. This talk will review main trends in the area of distributed inference for wireless sensor networks, including message passing, consensus, and distributed optimization, and it will also present a set of tools that are used in the design and analysis of such algorithms. A special emphasis will be given to the methodology of large deviations, which proved to be a powerful tool in the design of distributed inference algorithms.
Bio: Dragana Bajović received a dipl. ing. degree from the School of Electrical Engineering, University of Belgrade, Serbia, in August 2007, and the PhD degree in electrical and computer engineering from Carnegie Mellon University, Pittsburgh, PA, and Instituto de Sistemas e Robótica (ISR), Instituto Superior Técnico (IST), Lisbon, Portugal, in May 2013. Since October 2013, she has been a research fellow at the BioSense Center, University of Novi Sad, Serbia. From June to September 2013, she was a postdoctoral researcher at IST. Her research interests include statistical signal processing and large deviations analysis for sensor networks.
Multiple vehicle motion planning: an infinite dimension Newton optimization method
July 10, 11h00-13h00
Abstract: We describe a numerical algorithm for multiple vehicle motion planning that addresses explicitly temporal and spatial specifications, as well as energy-related constraints. As a motivating example, we cite the case where a group of vehicles is tasked to reach a number of target points at the same time (simultaneous arrival problem) and avoid inter-vehicle as well as vehicle/obstacle collision, subject to the constraint that the overall energy required for vehicle motion be minimized.
The methodology adopted builds on a numerical method for solving optimal control problems that is known as the PRojection Operator based Newton method for Trajectory Optimization (PRONTO)---a method that avoids the transcription phase typical in direct methods for numerical optimal control and that employs an infinite dimension Newton method to achieve second order convergence of the trajectory optimization problem.
With the theoretical set-up adopted, the vehicle dynamics are taken explicitly into account at the planning level. Thus, in contrast to some of the planning methods available in the literature, the method proposed allows for the direct incorporation of dynamical constraints imposed by the physical characteristics of the vehicles, motion actuators, and even energy sources (e.g. batteries). Should the problem to be solved be feasible, the method yields energy-optimal trajectories without the need to separate the steps of path planning and trajectory generation, as is customary in many of the motion planning methods described in the literature. Restrictive system properties such as differential flatness are not required.
We formulate the problem of multiple vehicle motion planning, describe the theoretical procedure adopted to solve it, and detail the implementation of the corresponding numerical algorithm. The efficacy of the method is illustrated through examples of the simultaneous arrival problem for the realistic case where the initial estimates of the trajectories, that is, the seeds for the optimization procedure are obtained from simple geometrical considerations; for example, by joining the initial and final target points of the vehicles via straight lines. Even though the thereby obtained initial trajectories yield inter-vehicle and vehicle/obstacle collisions, in the event that the optimization problem is feasible, the projection operator algorithm will generate appropriate trajectories that minimize the overall vehicle energy spent and meet the required temporal and spatial constraints.
Bio: Andreas J. Häusler studied computer science with a minor in logic and theory of science at the Ludwig-Maximilians-Universität (LMU) in Munich. After obtaining a joint M.Sc. degree in cognitive robotics from the LMU and the Technical University of Munich (TUM) in September 2006, he became a member of three research groups, working in cognitive robotics, artificial intelligence, and the semantic web. In August 2007, he joined the Institute for Systems and Robotics (ISR) of the Instituto Superior Técnico (IST) in Lisbon, Portugal, as a Marie Curie Early Stage Researcher at the Dynamical Systems and Ocean Robotics laboratory (DSOR). Since January 2011, he is working towards the Ph.D. degree in Marine Robotics. His research interests include multiple vehicle trajectory generation, optimization, cooperative planning, and propeller theory.
Cooperative Robot Localization and Target Tracking Based on Least Squares Minimization
July 10, 8h30-10h30
Abstract: In this work we address the problem of cooperative localization and target tracking with a team of moving robots. We model the problem as a least squares minimization problem and show that this problem can be efficiently solved using sparse optimization methods. To achieve this, we represent the problem as a graph, where the nodes are robot and target poses at individual timesteps, and the edges are their relative measurements. Static landmarks at known position are used to define a common reference frame for the robots and the targets. In this way, we mitigate the risk of using measurements and state estimates more than once, since all the relative measurements are i.i.d. and no marginalization is performed. Experiments performed using a set of real robots show higher accuracy compared to an Extended Kalman filter.
Bio: Aamir Ahmad received his Ph.D. degree (with merit and European Doctorate) in Electrical and Computer Engineering from Instituto Superior Técnico (IST), Lisbon, Portugal in April, 2013. He received his Bachelors degree, B-Tech. (with Honors), in Civil Engineering from the Indian Institute of Technology (IIT) Kharagpur, India in July, 2008. From 2013 onwards, he is a postdoctoral researcher at the Institute for Systems and Robotics (ISR), IST, Lisbon. From 2009 to 2013 he was a Ph.D. student at the same place where he was being advised by Prof. Dr. Pedro U. Lima. His main research focuses on the fields of sensor fusion, multi-robot systems, robot localization, object tracking, cooperative methods for tracking and localization, object detection, 3D recognition and optimization techniques for localization and tracking. he served as a program committee member of the international conference on autonomous agents and multiagent systems (AAMAS) 2014 and the European Conference on Mobile Robotics (ECMR) 2013. He is also an organizing committee member of the Middle Sized League in RoboCup 2014, João Pessoa, Brazil.
Multi-agent Decision Making Under Uncertainty: Theory and Practical Applications
July 9, 14h00-16h00
Abstract: Decision-Theoretic mathematical frameworks based on Markov Decision Processes (MDPs) provide a principled way to model and solve problems of sequential decision-making under uncertainty. In this talk, I will focus on the topic of cooperative multiagent decision-making, for which there is a large family of MDP-based frameworks. These frameworks differ in their assumptions regarding the possible interactions between different agents, either through their actions or through communication, and consequently they are targeted towards different applications. I will discuss the fundamentals of the most popular frameworks in this class: from fully-communicative (centralized) frameworks (MMDPs, MPOMDPs), to fully decentralized decision-making models, such as Dec-MDPs and Dec-POMDPs; interactive POMDPs (I-POMDPs) in which agents reason over models of the properties, behavior and intentions of other agents; and scalable models that make structural assumptions on the decision-making problem (TD-POMDPs, ND-POMDPs).
Since these frameworks provide abstract mathematical models of decision-making, where the dynamics of the underlying system are often simplified in exchange for computational tractability, there are non-trivial issues regarding the implementation of the respective solutions to the control of physical systems. I will describe the most notable applications of decision-theory to the control of multi-robot and networked robot systems, and discuss how the level of abstraction of the decision-making problem can influence the complexity and the validity of its Markovian models. Finally, I will present recent methods for event-driven decision-making under uncertainty that have been successfully applied in the context of multi-robot autonomous surveillance.
Bio: João Messias is a post-doctoral researcher at ISR-Lisbon. He received his PhD at Instituto Superior Técnico (IST) in 2014, under the supervision of Pedro Lima (ISR-IST) and Matthijs Spaan (TU Delft), on the topic of decision-making under uncertainty for multi-robot systems. He is currently working on long-term applications of networked robot systems for human assistance. His research interests include Cooperative Robotics, planning under uncertainty, and mobile robot navigation. João has participated and contributed to AAMAS/AAAI since 2010, and has served on the program and review committee of several related conferences. He has also been a regular participant/team leader in international robotics competitions (RoboCup) since 2009. João is a co-organizer for the 2014 Workshop on Multiagent Sequential Decision Making Under Uncertainty (MSDM).
Sequential Decision-Making Under Uncertainty: an Introduction (MDPs, POMDPs, Reinforcement Learning)
July 9, 9h00-12h00
Abstract: Multi-agent and (Multi-)Robot systems need systematic approaches to task modeling and planning, supported by formal methods, so as to enable stating performance bounds and several properties of a plan to carry out a task. Such formal approaches are crucial to ensure practical (often non-intuitive) results when applied to real systems, that scale up well with the several dimensions (e.g., number of robots/agents, state space size) of the problem at hand. In this module I will provide an introduction to the fundamental concepts from sequential decision-making under uncertainty, that pave the way for subsequent talks on multi-agent decision-making under uncertainty and cooperative perception. These are approaches based on Bayesian probability theory, and which determine optimal plans which are maps from the states (or the belief about the states) where the agent/robot is onto the the available actions for the robot/agent. Markov Decision processes (MDPs), their Reinforcement Learning solution, and Partially Observable MDPs (POMDPs) will be covered by the lecture.
Bio: Pedro Lima got his Electrical Engineering degree at Instituto Superior Técnico in 1984, where he also got a MSc degree in 1989. Ph.D. (1994) in Electrical Engineering at RPI, NY, USA. Currently, he is a Professor at Instituto Superior Técnico, Universidade de Lisboa, and a researcher of the Institute for Systems and Robotics, where he is the coordinator of the Intelligent Robots and Systems group and Deputy President for Scientific Affairs. He is the co-author of two books, and member of the Editorial Board of the Elsevier’s Journal of Robotics and Autonomous Systems. His research interests lie in the areas of discrete event models of robot tasks and planning under uncertainty, with applications to networked robot systems.
Pedro Lima was a Trustee of the RoboCup Federation (2003-2012), and was the General Chair of RoboCup2004, held in Lisbon.
He has also been very active in the promotion of Science and Technology to the society, through the organization of Robotics events in Portugal, including the Portuguese Robotics Open since 2001.
Cooperative Navigation and Control of Networked Autonomous Vehicles
July 8, 15h00-18h00
Abstract: The last decade has witnessed tremendous progress in the development of marine technologies that are steadily affording scientists advanced equipment and methods for ocean exploration and exploitation. Recent advances in marine robotics, sensors, computers, communications, and information systems are being applied to the development of sophisticated technologies that will lead to safer, faster, and far more efficient ways of exploring the ocean frontier, especially in hazardous conditions. As part of this trend, there has been a surge of interest worldwide in the development of autonomous marine robots capable of roaming the oceans freely and collecting data at the surface of the ocean and underwater on an unprecedented scale. Representative examples are autonomous surface craft (ASC) and autonomous underwater vehicles (AUVs). The mission scenarios envisioned call for the control of single or multiple AUVs acting in cooperation to execute challenging tasks without close supervision of human operators.
This talk addresses the general topic of cooperative navigation and motion control of marine vehicles both from a theoretical and a practical standpoint. The presentation is rooted in practical developments and experiments. Examples of scientific mission scenarios with ASCs and AUVs, acting alone or in cooperation, set the stage for the main contents of the presentation. From a theoretical standpoint, special attention is given to a number of challenging problems that include: i) cooperative motion control of fleets of autonomous vehicles, ii) optimal sensor placement for multiple underwater vehicle localization with acoustic range measurements, and ii) cooperative vehicle navigation using geophysical and single-beacon measurements. The results obtained are illustrated with videos from actual field tests with multiple marine robots. The core material presented in the talk was obtained in the scope of the GREX (http://www.grex-project.eu), CO3AUVs (http://www.co3-auvs.org), and MORPH (http://morph-project.eu/) projects of the EC.
Bio: António M. Pascoal received the Licenciatura degree in Electrical Engineering from the Instituto Superior Técnico (IST), Lisbon, Portugal in 1974, the M.S. degree in Electrical Engineering from the University of Minnesota, Minneapolis, Minnesota, USA in 1983, and the Ph.D. degree in Control Science from the same school in 1987. From 1987-88 he was a Research Scientist with Integrated Systems Incorporated, Santa Clara, California, where he conducted research and development work in the areas of system modeling and identification and robust and adaptive control. Since 1998 he has been with the Department of Electrical Engineering of IST, where he is currently an Associate Professor of Control and Robotics. He has coordinated and participated in a large number of international projects that have led to the design, development, and field-testing of single and multiple autonomous marine and air vehicles.
His research interests include linear and nonlinear control theory, robust adaptive control, and networked planning, navigation, and control of multiple autonomous vehicles with applications to air, land, and underwater robots. His long-term goal is to contribute to the development of advanced robotic systems for ocean exploration and exploitation.
Data Science: Rethinking DSP
July 8, 9h00-13h00
Abstract: We provide a brief introduction and background to the challenges and opportunities that arise in Network Science and Data Science. Then we focus on the data deluge. Chris Anderson titled provocatively his 06.23.08 piece “The End of Theory: The Data Deluge Makes the Science Method Obsolete.” Data (big), computers (cloud), storage (vast), bandwidth (massive) and Google (or the likes) will find the correlations that will save the day. No (need for) causation. May be; or we might still try to explain it. Data is big, comes from all sorts of sources – social, business, urban, physical, biological, molecular, to name a few. However, if we do capture the relations among data through (arbitrary) graphs (and this in itself is a big if), the “big data” challenge can be cast in the familiar setting of everyone’s beloved DSP. This talk will overview our progress so far extending to data defined on graphs, graph signals, traditional signal processing concepts including shifting, frequency, filtering, convolution, spectral representation, filters frequency response, linear transforms like the discrete Fourier transform. We illustrate with data drawn from social networks and the World Wide Web.
Bio: José M. F. Moura is a visiting Professor at CUSP, NYU (2013-14). He is the Philip and Marsha Dowd University Professor at CMU He was a Professor at IST and a visiting Professor at MIT. His interests are in signal processing (SP) and data science. A sequence detector co-invented with Kavcic (US patents `839 and `180) is found in 2.4 billion disk drives (60% of the computers sold worldwide in the last 10 years). He cofounded SpiralGen to commercialize Spiral (www.spiral.net) (CMU license). He was an IEEE Board Director, President of the IEEE SP Society (SPS), and Editor in Chief for the Transactions on SP. Moura received the IEEE SPS Technical Achievement Award and the IEEE SPS Society Award for outstanding technical contributions and leadership in SP. He is an IEEE Fellow, AAAS Fellow, Academy of Sciences of Portugal corresponding member, US National Academy of Engineering member.