The Max Planck Institute for Intelligent Systems currently houses multiple Research Groups, all of which are lead by outstanding scientists who receive a secure budget. This freedom helps them propel their research and lay the foundations for a successful scientific career.
The Autonomous Learning group’s mission is to make robots learn in a way that is similar to early childhood development. Learning directly from interacting with their environment can potentially enable them to act in a complex and constantly changing world, which they cannot be programmed to do.
The Autonomous Vision research group, which is based at the Max Planck Institute for Intelligent Systems in Tübingen and the University of Tübingen, addresses questions related to robustness as well as methods that enable high-capacity models (such as deep neuronal networks) to learn with a small amount of data. More specifically, the group’s research focuses on robust perception for autonomous agents, especially autonomous vehicles. Research activities range from sensor-based perception (3D reconstruction, motion estimation, object recognition) and holistic scene interpretation (3D lane and intersection estimation), to sensor engine control approaches.
Animals not only run dynamically, efficiently, and elegantly, they also quickly adapt to new terrain while moving on it. Their locomotion is a carefully orchestrated interplay of muscles and tendons that has been optimized over the course of evolution. Alexander Badri-Spröwitz and his team use robots and simulations to understand animals and their movements. They investigate why an animal activates a muscle, what forces enable the animal to move or why not all muscles and tendons are the same. The researchers take inspiration from animals to build robot models. With robots, researchers can directly test the function of the individual parts. Their findings could help to improve walking robots, prostheses or exoskeleton technologies – external support structures for the body.
Intelligent systems such as robots need the ability to learn and adapt to their environment. The Embodied Vision research group investigates novel methods that make it possible to understand dynamic 3D scenes. Such knowledge is required for artificial intelligent systems to solve complex tasks such as autonomous navigation or object manipulation.
Most common approaches that enable robots to carry out specific tasks are based on different components for perception and control. In contrast, the group’s researchers develop holistic methods that allow robots to learn how to interact with the environment and perform tasks. To do this, the robots should learn to draw on the sensor data of cameras and touch sensors to develop a model of their surroundings that is related to their task. They should also learn to predict the immediate effects of their actions. In turn, the knowledge they acquire from the model could enable robots to choose their actions themselves.
Intelligent Control Systems
The Intelligent Control Systems (ICS) group focuses on fundamental questions of future intelligent systems, which will be able to autonomously interact with their environment by perceiving the world, acting according to a goal, and learning from both. For instance, the group’s researchers investigate how a machine can independently learn new tasks from data reliably, safely, and efficiently. They also look at how collectives of several intelligent systems can carry out a task together – such as several robots coordinating their motion or autonomous vehicles driving in a convoy. It all comes down to sophisticated decision-making and learning algorithms, which are essentially the intelligent system’s brain.
Starting with mathematical problem descriptions and analysis, the team develops new algorithms and methods that can be applied in many different future systems. Going beyond models and simulations, the team validates its research in laboratory experiments, for instance on a humanoid robot learning to balance a stick in its hand, or multiple dynamical systems coordinating their motion over large-scale wireless networks. They also implement their algorithms outside the lab with Cyber Valley industry partners.
Research at the ICS group is interdisciplinary and spans engineering, computer science, mathematics, and machine learning. The main research directions are currently learning-based control, distributed and networked systems, and resource-efficient algorithms.
Locomotion in Biorobotic and Somatic Systems
Led by Ardian Jusufi, the scientists of the independent Cyber Valley “Locomotion in Biorobotic and Somatic Systems” research group investigates animals’ locomotion and physical structure, as perfected by nature. The researchers then apply their biological findings to the development of life-like robots. Their research is at the interface between engineering and biology – a relatively new and promising field.
Soft robotics is one of the group’s fields of research. While most of today’s robots are still made of hard, rigid components, soft robotics aims to incorporate flexible, malleable components into synthetic systems. To this end, scientists draw upon the morphological intelligence that is found in the body structure of mammals, insects or reptiles which allows them to move efficiently and robustly. Based on these natural design principles, they develop robots that can walk, run, or swim like their animal counterparts. Of particular interest are the interactions of stronger or stiffer tissues with softer or more flexible ones; animal locomotion is a perfectly coordinated action of all materials. By incorporating morphological intelligence into swimming or climbing robots, Jusufi and his team are confident that such machines can cope better with complex environments and overcome obstacles more easily.
However, the scientists’ motivation goes even further; when building such robots, the researchers gain insights into the animal, how its locomotion is adapted to its terrain, and they can answer questions as to why evolution has produced certain structures. Ultimately, the scientists aim to gain a deeper understanding of both animals and machines simultaneously.
Micro, Nano, and Molecular Systems
Headed by Professor Peer Fischer, the independent Max Planck Research Group “Micro, Nano and Molecular Systems” investigates the physical and chemical properties of active matter, develops unique nanofabrication methods, and builds nanorobotic systems that are smaller than a human cell. For example, the group is working on microswimmers, nanomotors driven by chemical reactions, and nanopropellers that can move through biological tissue. Such “nanobots” hold great potential in applications such as the minimally invasive medical technology of the future. The group has also developed a state-of-the-art vapor deposition technique with which it can produce hundreds of billions of nanostructures quickly and with high precision. Moreover, the group has invented the acoustic hologram, with which ultrasound waves can be formed threedimensionally and the most precise ultrasound fields to date can be generated.
Movement Generation and Control
Complex action sequences, such as running on uneven terrain, are relatively easy for adults, but still present great challenges for robots. Led by Dr. Ludovic Righetti, The Motion Generation and Control research group investigates algorithmic principles that allow humanoid robots to perform complex motion sequences, such as lifting a cushion while simultaneously grasping an object underneath it. The research aims to develop basic principles of movement and manipulation that robots need to adapt to unknown variables and changing environments, and thus act efficiently and autonomously.
In order to be reliable and safe, robots must be able to react quickly to unpredictable events. The research group’s scientists have developed an algorithm that can calculate optimal movements in less than a second – making it the fastest to date. This has made it possible to develop novel methods for robotics that make physical contact between robots and their environment significantly more stable.
In cooperation with the MPI-IS “Dynamic Locomotion” research group, Righetti’s team created the world’s lightest force-controlled four-legged robot. This robot’s movement is very dynamic, which makes it an ideal platform for evaluating new control and learning algorithms. With this four-legged animal, researchers can test the capabilities of their motion planning algorithm for walking and jumping tasks.These movements are very robust to external disturbances, even when the robot is pushed or the ground moves.
The research team is also investigating how robots can learn from previous experiences, both positive and negative. In collaboration with other research teams, the Motion Generation and Control group has developed a new exploration strategy based on machine learning methods. This strategy makes it possible to generalize and optimize movements toward unknown objects or accessible places.
Physics for Inference and Optimization
Caterina De Bacco
The Physics for Inference and Optimization Group’s research focuses on understanding relations between the microscopic and macroscopic properties of complex large-scale interacting systems, such as networks. In cooperation with experts from other disciplines, De Bacco and her team develop models and algorithms based on principles of statistical physics. This knowledge could be used, for instance, to modify the interactions of a network’s individual constituents and thus to optimize its overall properties.
One of the group’s research interests is routing optimization. By collecting data on individual drivers and the vehicles in their immediate surroundings, researchers can draw conclusions on the behavioral patterns of this small group, and use these to make a generalization about all driving behaviors. In other words, they zoom in on part of the whole, closely observe behavioral patterns, and project their findings onto the big picture. Such knowledge can be used to calculate the best possible individual routes for all drivers by optimizing traffic management, even if this may mean a longer distance for the individual. In a collaboration with the Mathematics Department at the University of Padova, De Bacco’s group developed an efficient algorithm capable of deriving optimal general solutions for many routing problems.
De Bacco and her team also focus on investigating inference problems on networks. Inference aims to estimate the parameters of a model that is believed to have generated certain data. De Bacco investigates, for example, how likely it is that members of a social network will interact with one another. In this research area, for instance, she and her group recently developed a model that serves to estimate a node’s measure of importance in a network (known as eigenvector centrality) from a graph sample. This measure is particularly relevant when retrieving the information for the whole network is not feasible, as is the case with social networks.
Probabilistic Learning Group
Valera and her group focus on developing machine learning methods that are flexible, robust, and fair. Flexible means they are capable of modeling complex real-world data, which are often heterogeneous in nature and collected over time. Secondly, the group’s research aims to improve the robustness of algorithms. An algorithm is considered robust when it is able to point out “what it does not know”. This means that, in addition to making predictions, it also expresses how probable they are. The group’s research findings have given rise to new methods and software applications for automatic data pre-processing.
In addition, Valera and her team are researching ways to make algorithms that are part of important decision-making processes fairer. To this end, the scientists have worked on translating legal definitions of fairness into a mathematical formula. Valera and her team are thus designing new algorithms that are both accurate and fair, and which can be quantified with the support of observational data.
The group’s researchers have already achieved significant results in the field of fairness and machine learning. Topics of the group’s publications have included approaches to improving the performance and fairness of human decision making, as well as an efficient learning procedure to transform an original (unfair) classifier into a fair one.
The group’s research can be applied in a broad range of fields, from medicine and psychiatry to fields that are socially relevant. Valera’s research thus pays close attention to the ethical challenges of decisions that are supported by algorithms and could have far-reaching consequences for the people concerned. These issues play a major role in areas such as hiring or loan approval processes.
Numerical Problems --- linear algebra and optimization, integration and the solution of differential equations --- are the computational bottleneck of artificial intelligent systems. Intriguingly, the numerical algorithms used for these tasks are also compact little intelligent agents themselves. They estimate unknown / uncomputable quantities by observing the result of feasible computations. They also actively decide which computations to perform.
The Research Group on Probabilistic Numerics studies this philosophical and mathematical connection between computation and inference. We aim to build a theoretical understanding of numerical computer algorithms as agents acting rationally under uncertainty. We analyse existing algorithms from this viewpoint, and propose novel algorithms that provide functionality for key computational challenges in the science of Intelligent Systems.
The Max Planck Research Group for Rationality Enhancement aims to develop a scientific foundation and intelligent technologies that enable people to become more effective. To this end, its research focuses on cognitive growth, goal setting, and goal achievement. The group elucidates the underlying mechanisms and investigates how they can be promoted, supported, and improved. If it succeeds, the group’s findings will revolutionize self-improvement, personal development, psychiatry and psychotherapy, brain training, and education.
Statistical Learning Theory
Ulrike von Luxburg
The goal of statistical learning theory is to provide a solid mathematical basis for machine learning algorithms and to analyze their behavior. The scientists of the Max Planck Fellow Group aim to assess whether the results achieved by machine learning algorithms are trustworthy, whether the algorithms work or not, or how complex they are in terms of data required or computation time needed.
The group’s researchers focus on the area of comparison-based machine learning, which is a subfield of machine learning. The researchers consider a particular scenario where the input to a machine learning algorithm can be collected in a human-friendly way. They consider a setting where the input to a machine learning algorithm is not given in terms of similarity values (“On a scale from 0 to 1, the similarity between image A and image B is 0.8”), but rather in terms of distance comparisons (“Image A is more similar to image B than to image C”). Many studies in psychology show that, for people, such qualitative comparisons are much easier to provide than quantitative similarity scores.