Learning and control for complex and multi-agent systems
Event details
| Date | 17.02.2021 |
| Hour | 09:00 › 10:00 |
| Speaker | Dr. Yvonne Stürz, Model Predictive Control Laboratory, UC Berkeley |
| Location | Online |
| Category | Conferences - Seminars |
Abstract: The increased availability of inexpensive hardware, sensors and computational power enables a widespread use of systems in completely novel application areas and with ever-growing complexity. While automation and multi-agent systems have a significant potential to increase the efficiency, sustainability and safety in many industries (such as energy, traffic, robotics, construction), there are still major open challenges in the control of complex and multi-agent systems. In particular, their large scale and spatial distribution give rise to the need of efficient and scalable local control algorithms.
Moreover, safety and performance have to be guaranteed despite uncertainties and unknown environments. The talk will illustrate this need for novel advanced control methods with two examples:
The first part of the talk introduces a novel control application in the area of digital fabrication. Unlike many other branches of industry, the construction sector has not followed the trend of automation, mainly because it has to deal with unstructured environments and non-standardized complex tasks. However, as one of the largest consumers of resources worldwide, the construction industry has a high potential for improved sustainability. A novel method to precisely control the form of a tensioned cable-net based formwork on site is an example of how this potential can be achieved. The presented approach enables the construction of lightweight building elements. The developed optimization-based control algorithm tackles the complexity of the flexible cable-net system by exploiting the mathematical structure and convexity of the problem. Experimental validation on a cable net for a doubly-curved roof shell will be presented.
The second part of the talk focuses on data-driven control of autonomous multi-agent systems. While the operation of single robots in structured environments, such as autonomous driving on the highway, is well studied, the coordination of complex tasks of fleets of multiple agents still poses major challenges. In particular, designing distributed controllers which guarantee safety and a high performance for multi-agent systems is difficult. This is due to the inherent requirements for scalability and the thereby introduced limitations in local sensing, computational power and communication. These informational constraints can lead to a decreased control performance. Therefore, a novel distributed learning model predictive control scheme is presented, which uses local data in order to learn local control policies. This approach is data-efficient and guarantees safety as well as, under mild assumptions, global optimality, of the centralized closed-loop system.
Bio: Yvonne Stürz is a postdoctoral researcher at the Model Predictive Control Laboratory at UC Berkeley, USA. Through the Marie-Curie fellowship that she was awarded, she is also affiliated with the Automatic Control Department at KTH, Sweden. In 2019, she obtained her PhD from the Automatic Control Laboratory at ETH Zurich, Switzerland. She holds a Bachelor's and Master's degree in mechanical engineering from TU Munich, and a Diploma degree in general engineering from CentraleSupelec Paris, France. She was awarded scholarships from the Hans Werthen Foundation, the German National Academic Foundation and the Ulderup Foundation. She gained industrial experience through internships with Bosch, Porsche and Siemens in Germany and with Siemens SFAE in China. Her research interests include distributed optimal control and data-driven methods, with applications to large-scale heterogeneous interconnected systems, multi-robot systems, autonomous driving and digital fabrication.
Moreover, safety and performance have to be guaranteed despite uncertainties and unknown environments. The talk will illustrate this need for novel advanced control methods with two examples:
The first part of the talk introduces a novel control application in the area of digital fabrication. Unlike many other branches of industry, the construction sector has not followed the trend of automation, mainly because it has to deal with unstructured environments and non-standardized complex tasks. However, as one of the largest consumers of resources worldwide, the construction industry has a high potential for improved sustainability. A novel method to precisely control the form of a tensioned cable-net based formwork on site is an example of how this potential can be achieved. The presented approach enables the construction of lightweight building elements. The developed optimization-based control algorithm tackles the complexity of the flexible cable-net system by exploiting the mathematical structure and convexity of the problem. Experimental validation on a cable net for a doubly-curved roof shell will be presented.
The second part of the talk focuses on data-driven control of autonomous multi-agent systems. While the operation of single robots in structured environments, such as autonomous driving on the highway, is well studied, the coordination of complex tasks of fleets of multiple agents still poses major challenges. In particular, designing distributed controllers which guarantee safety and a high performance for multi-agent systems is difficult. This is due to the inherent requirements for scalability and the thereby introduced limitations in local sensing, computational power and communication. These informational constraints can lead to a decreased control performance. Therefore, a novel distributed learning model predictive control scheme is presented, which uses local data in order to learn local control policies. This approach is data-efficient and guarantees safety as well as, under mild assumptions, global optimality, of the centralized closed-loop system.
Bio: Yvonne Stürz is a postdoctoral researcher at the Model Predictive Control Laboratory at UC Berkeley, USA. Through the Marie-Curie fellowship that she was awarded, she is also affiliated with the Automatic Control Department at KTH, Sweden. In 2019, she obtained her PhD from the Automatic Control Laboratory at ETH Zurich, Switzerland. She holds a Bachelor's and Master's degree in mechanical engineering from TU Munich, and a Diploma degree in general engineering from CentraleSupelec Paris, France. She was awarded scholarships from the Hans Werthen Foundation, the German National Academic Foundation and the Ulderup Foundation. She gained industrial experience through internships with Bosch, Porsche and Siemens in Germany and with Siemens SFAE in China. Her research interests include distributed optimal control and data-driven methods, with applications to large-scale heterogeneous interconnected systems, multi-robot systems, autonomous driving and digital fabrication.
Practical information
- General public
- Free