2019 Intelligent Systems Summer Colloquium

Biological Design and Integrative Structures for Architecture
In biology most load-bearing structures are fiber composites. They are made from fibers, as for example cellulose, chitin or collagen, and a matrix material that supports them and maintains their relative position. The astounding performance and unrivalled resource efficiency of biological structures stems from these fibrous systems. Their organization, directionality and density is finely tuned and locally varied in order to ensure that material is only placed where it is needed. Our projects aim to transfer this biological principle of load-adapted and thus highly differentiated fiber composite systems into architecture. Manmade composites, such as the glass- or carbon-fiber-reinforced plastics are ideally suited for such an approach, because they share their fundamental characteristics with natural composites. Many years of biomimetic research at the Institute for Computational Design and Construction (ICD) and the Institute for Building Structures and Structural Design (ITKE) shows how an interdisciplinary exploration of biological principles together with robotic fabrication strategies can lead to truly novel and genuinely building systems. The research on fiber composite building systems and related computational design and robotic fabrication methods will be further pursued in the context of the new Cluster of Excellence “Integrative Computational Design and Construction for Architecture” at the University of Stuttgart.

Jan Knippers is a practicing consulting engineer and since 2000 head of the Institute for Building Structures and Structural Design (ITKE) at the University of Stuttgart. His interest is in innovative and resource-efficient structures created at the intersection of research and development and practice.
 
From 2014 to 2019 Jan Knippers was the coordinator of the DFG collaborative research center Biological Design and Integrative Structures. Since 2019 he is Deputy Executive Director of the Cluster of Excellence Integrative Computational Design and Construction for Architecture and Vice-rector for Research of the University of Stuttgart. He is author of several books, numerous scientific publications, and member of various advisory boards.

Machine learning for behavior generation and haptic sensation
Georg Martius is interested in making robots learn like children. In his talk, he will speak about their recent advances in making robots learn from scratch. This involves guiding the exploration process to natural behaviors, learning model of the environment and planning. For a self-learning robot the physical interaction is very important. However, haptic sensation technology is currently bulky, fragile and expensive. Georg Martius will present a way to use machine learning to create a cheap and large surface haptic force-measurement device.

Georg Martius is leading a research group on Autonomous Learning at the Max Planck Institute for Intelligent Systems in Tübingen. Before joining the MPI in Tübingen, he was a postdoc fellow at the IST Austria in the groups of Christoph Lampert and Cašper Tkacik after being a postdoc at the Max Planck Institute for Mathematics in the Sciences in Leipzig. He pursues research in autonomous learning, that is how an embodied agent can determine what to learn, how to learn, and how to judge the learning success. He is using information theory and dynamical systems theory to formulate generic intrinsic motivations that lead to coherent behavior exploration – much like playful behavior. Together with Ralf Der he published a book on this topic [1]. With his research group he is also working on machine learning methods particularly suitable for internal models, reinforcement learning and haptics.
 
[1] playfulmachines.com

Deep Learning on Unstructured Point Clouds
Traditional convolution layers in deep learning of image and video data are specifically designed to exploit the natural data representation of images - a fixed and regular grid. However, unstructured data such as 3D point clouds contain irregular neighborhoods. A first challenge is to quickly establish the neighborhood for all points at the same time. In order to build k nearest neighbor graphs from scratch we present different massively parallel algorithms that are particularly suited for small and high dimensional point data. Furthermore, we introduce a natural generalization flex-convolution of the conventional convolution layer along with an efficient GPU implementation, which allows for transferring best-practices and design choices from 2D-image learning methods directly to point cloud processing. We demonstrate competitive performance on benchmark sets using fewer parameters and lower memory consumption and obtain significant improvements on a million-scale real-world dataset. Ours is the first which allows to efficiently process up to 18 million points concurrently.

Hendrik P. A. Lensch holds the chair for computer graphics at Tübingen University and is currently the head of the computer science department and the vice-spokesperson of the International Max Planck Research School for Intelligent Systems. He received his diploma in computers science from the University of Erlangen in 1999. He worked as a research associate at the computer graphics group at the Max-Planck-Institut für Informatik in Saarbrücken, Germany, and received his PhD from Saarland University in 2003. Hendrik Lensch spent two years (2004-2006) as a visiting assistant professor at Stanford University, USA, followed by a stay at the MPI Informatik as the head of an independent research group. From 2009 to 2011 he was a full professor at the Institute for Media Informatics at Ulm University, Germany. In his career, he received the Eurographics Young Researcher Award 2005, was awarded an Emmy-Noether-Fellowship by the German Research Foundation (DFG) in 2007 and received an NVIDIA Professor Partnership Award in 2010. His research interests include 3D appearance acquisition, computational photography, machine learning, global illumination and image-based rendering, and massively parallel programming.

Holograms for Acoustics
Ultrasound is widely known from its use in medical imaging. Apart from imaging, ultrasound has many emerging applications in contact-free manipulation and power transfer. These benefit from a precise control over its pressure distribution in space. Typically, ultrasound fields are generated with single piezo-transducers and then formed by a lens, horn, or resonator. Alternately, the output from multiple transducers is combined in a phased array configuration to obtain more complex fields. Such arrays are employed in medical imaging but require elaborate control electronics. In this talk I will present my work on the acoustic hologram — a new way of forming arbitrary sound fields using a phase plate and a single transducer element. Using digital holography and a 3D printer to fabricate the holograms, this simple method encodes fields with several orders of magnitude more degrees of freedom compared to established methods. The resulting complex sound fields have been shown to advance parallel particle assembly and fabrication, propel objects along trajectories and levitate matter in air. I will discuss how one can generate and image sophisticated ultrasound fields using the acoustic hologram, as well as applications that are possible with ultrasound-matter interactions.

Kai Melde received his engineering degree (Dipl.-Ing.) in mechatronics from the Technical University Dresden in 2009 focusing on microsystems design. From 2008 to 2013 he was a member of the technical staff at the Palo Alto Research Center (formerly Xerox PARC) in Palo Alto, California, where he worked on a wide range of research projects including inkjet printheads and water treatment. Since 2013 he is a PhD student in the Micro, Nano and Molecular Systems lab at the Max Planck Institute for Intelligent Systems.