INRIA, France

Geometric and end-to-end robot vision-based control

As for humans and most animals, vision is a crucial sense for a robot to interact within its environment. Vision-based robot motion control, also named visual servoing, is a general approach to close the perception-action loop. The aim of this talk is to provide a comprehensive state of the art on the basic concepts, methodologies, and applications. 

In a first part, the traditional approach based on geometric visual features, such as image points, image moments, or camera-object pose will be described. The more recent end-to-end approach that directly uses the image content without any image tracking nor matching process will be also considered, providing a link to CNN modern methods that use the same inputs. 

Biography

François Chaumette is an Inria senior research scientist at Irisa in Rennes. He received a PhD in Computer Science from the University of Rennes 1 in 1990. His research interests include robotics and computer vision, especially visual servoing and active perception. He supervised more than 30 PhD students and published over 300 journal or conference papers, with the 2002 Best IEEE Transactions on Robotics and Automation Paper Award, the 2019 Best IEEE Robotics and Automation Letters Paper Award, and the 2020 Best IEEE Robotics and Automation Magazine Paper Award. He was Founding Senior Editor of the IEEE Robotics and Automation Letters, member of the Editorial Board of the Int. Journal of Robotics Research, program co-chair of the IEEE Int. Conf. on Robotics and Automation in 2020, and served recently as Senior Editor of the IEEE Transactions on Robotics. He was a panel member for the ERC PE7 Consolidator grants in 2013, 2015, 2017 and 2019. He was recently involved in the H2020 EU Comanoid project. He is IEEE Fellow since 2013.

CTU Prague, Czech Republic

Training Neural Networks for Tasks with Non-Differentiable Objective Functions

Many computer vision tasks in their natural formulation have a non-differentiable objective function and the standard training procedure of a neural network is thus not applicable. Most deep learning methods side-step the problem by using a differentiable proxy loss, originally designed for another task, which may or may not align well with the non-differentiable objective. In the talk, we will present approaches for learning a differentiable surrogate of  decomposable and non-differentiable evaluation metrics.  For the decomposable case, the approach is validated on two practical tasks of scene text recognition and detection, where the surrogate learns an approximation of edit distance and intersection-over-union, respectively. For the non-decomposable case, we consider image retrieval and develop a  recall@k surrogate, also applicable for sorting and counting. Experiments confirm the superiority of the surrogates over proxy losses. 

Biography

Jiří Matas is a full professor and head of the Visual Recognition Group, Center for Machine Perception, CTU Prague. His research interests are in computer vision and pattern recognition. He has published more than 250 papers that have been cited more than 65,000 times and have an h-index of 93, according to Google Scholar. He has received several best paper awards at international conferences on computer vision. He is co-author of several international patents and co-founder of a spin-off and a start-up company. Prof. Matas has been actively involved as a partner in EU collaborative research projects and is also a project evaluator for the European Research Council and the EU Framework Program. He is also an editor of leading journals in the field of computer vision and pattern recognition. 

University of Maribor, Slovenia

Sensors and Other Micro-Photonics Structures Created at the Tip of an Optical Fiber

Optical fiber-based micro sensors, micro-photonics devices and methods for their manufacturing will be presented and discussed. Typical sensor devices have diameter of 0.125 mm, length not exceeding 1 mm, while being fusion spliced to the silica fiber tip. Creation of miniature and self-sustained sensors and photonics devices on the tip of an optical fiber can provide range of new approaches in industrial sensing, micro processing, and micro-opto-fluidic system design.  Smal size, all-silica design, cylindrical geometry, and possibility for their operation through optical fibers allows use of these devices in harsh environments, inaccessible locations, and space and weight restricted applications.

Biography

Denis Đonlagić received his MSc degree from the University of Maribor (Slovenia) in 1996 and the dual Ph.D. degrees from the University of Ljubljana (Slovenia) and University of Strathclyde, Glasgow (UK) in 1998 and 2000 respectively. He joined Corning Inc. (USA) during 2001 and 2002 as a Research/Development Scientist, where he worked in divisions of the Applied Fiber Research and Specialty Fiber Development. Since 2002 he has been leading the Laboratory for Electro-Optics and Sensor Systems at the University of Maribor (Slovenia), Faculty of EE and Comp. Sci., where he has been serving as a Full Professor since 2008, and Head of the Automation Department since 2010. Prof. Donlagic's research interests include Optical Fiber Sensors, Optical Fibers, Optical Fiber-based devices, MOEMS, and Opto-electronics System design, with a strong focus on applications of these fiber-based technologies. During his time at the university, he has been continually developing collaborations with leading global industrial partners in the field of Optical Fibers and Sensors. Prof. Donlagic also acted as a shareholder and technology adviser of Optacore d.o.o. between 2006 and 2017, a privately-owned SME dedicated to the production of specialty fibers and fiber manufacturing equipment. Optacore was acquired, and became part of Lumentum Inc. (USA) in 2017. He also serves as an Associate Editor for Applied Optics (OSA), as an Editorial Board member of Journal of Lightwave Technology and of Sensors (MDPI), and was a member of the Scientific Council of Slovenia’s Research Agency (2010-2015) and a Steering Board member of Slovenia’s Technology Agency (2006-2011). 

Eindhoven University of Technology, Netherlands

Energy Based Distributed Cooperative Model Predictive Control with Stability Guarantees 

This talk considers the design of distributed model predictive control (MPC) algorithms for networks of dissipative nonlinear systems, with the aim of achieving stability of the overall network. This research is motivated by frequency and voltage stability problems that arise in microgrids (or traditional power systems), but the results are applicable to general interconnected nonlinear systems. We first recall the state-of-the-art in compositional stability certificates for networks of dissipative nonlinear systems and we focus in detail on a specific set of relaxed conditions for stability, termed cyclically neutral supply conditions. Then we exploit the fact that real-life physical systems, such as generators in a microgrid, inherently satisfy an energy based dissipation inequality in terms of local storage and supply functions. We show that in order to achieve stability of the overall network it suffices to design local MPC controllers that co-operate to achieve cyclically neutral supply over the network. This type of distributed cooperative MPC (DC-MPC) solution is very attractive for controlling heterogenous microgrids, as it enables each system in the microgrid to contribute to the stabilization of the complete grid. Also, utilizing energy based storage and supply functions renders the design of the MPC controllers intuitive and physics-compliant. The complete methodology is illustrated for the case study of frequency stabilization in microgrids, which requires solving some specific challenges: dealing with time-varying equilibria and non-positive definite Bregman storage functions. We show that the developed DC-MPC scheme can outperform an alternative distributed averaging control algorithm (DAPI) for a benchmark microgrid example from the literature.  

Biography

Dr. Mircea Lazar is an Associate Professor in Constrained control of complex systems at the Electrical Engineering Department, Eindhoven University of Technology, The Netherlands. Lazar received the European Embedded Control Institute Ph.D. Award in 2007 for his PhD dissertation and a Veni personal grant from the Dutch Research Council (NWO) in 2008. He supervised 10 PhD researchers (2 received the Cum laude distinction) that received the PhD title. Lazar chaired the 4th IFAC Conference on Nonlinear Model Predictive Control, Noordwijkerhout, The Netherlands, in 2012. His research interests cover physics-based neural networks, nonlinear and data-driven predictive control, non-monotone Lyapunov functions, compositional stability certificates and distributed control. His research is driven by control problems in high-precision mechatronics, power electronics, power networks, water networks, automotive and biological systems. Lazar published 13 papers in IEEE Transactions on Automatic Control and 15 papers in Automatica. He is an Active Member of the IFAC Technical Committees 1.3 Discrete Event and Hybrid Systems, 2.3 Nonlinear Control Systems and an Associate Editor of IEEE Transactions on Automatic Control.