Research Statement: Humans, Robots & A.I.
My goal as a researcher is to advance the scientific and technological foundations for creating intelligent machines capable of autonomous interaction and learning in the real world. I believe this can only be achieved by co-designing embodiment and intelligence, jointly forming the driving force of advancement. For this, new systems, sensors, and materials need to be combined with the vast possibilities of control and machine learning algorithms, ever-increasing computing power, and high-speed network communication. The coming generations of machines will substantially increase our quality of life and work, collectively augmenting the human scientific intellect and experimentation bandwidth to accelerate knowledge discovery processes exponentially. Overall, such systems will form in diverse forms and functions the technological backbone in the globally and ubiquitously connected cyber-societies of the 21st century.
In my group, we bridge the disciplines of control, robotics, collective machine learning, and human motor intelligence. Our exploratory urge pushes the envelope in these disciplines and allows us to design the fundamental principles for engineering artificial intelligence and intelligent robots. Our robot assistants support the elderly in times of demographic change or human workers in the coming humancentered industrial revolution. We develop brain-controlled symbiotic prostheses and exoskeletons that allow people with disabilities to compensate for lost motor functionalities. The cooperative, networked laboratory helpers provide almost arbitrarily scalable experimentation to human scientists, allowing for unprecedented experiment-based knowledge discovery accelerations in molecule synthesis, drug synthesis, or cell analysis. The environmental intelligence formed with heterogeneous robot teams enables autonomous habitat exploration, monitoring, and intervention on land, water, and air.
In this Research Statement, I describe my scientific agenda and future work across distinct research fields that operationalize this vision:
- The Robot: dynamics and modeling, nonlinear control and estimation, tactile robots
- The Human: motor learning, injury protection and safety
- The Discovery: machine intelligence and autonomous scientific discovery
The Robot
Modeling and identification of intricate dynamics in complex mechanical and biomechanical systems for a variety of applications is at the core of understanding and building complex machines. They are central to any sophisticated control and learning algorithm for embodied AI systems such as industrial or service robots, drones, autonomous cars, or prostheses, as well as their digital twins. My group’s future work focuses on developing full-stack AI algorithms to autonomously generate morphological and dynamic models for arbitrary and reconfigurable robot systems from pure sensory perception only. The proposed algorithms build on methods from differential geometry and integrate first principles and information theory with model-informed machine learning.
Our complementary research stream on model-based nonlinear estimation and control for complex dynamical systems impacts a wide range of applications, including robotics, autonomous vehicles, and aerospace engineering. The developed methods allow for making more accurate predictions about system behavior, enabling optimal policies and robust controllers that can learn and adapt quickly when faced with unexpected changes, disturbances, or uncertainties in the real world. A central objective of our future work is to develop unified dexterous and tactile multi-fingered manipulation control and learning algorithms rooted in the energy-based passivity framework. Also, the modularization of such energy-based learning controllers and policies into networked - regardlessly stable - equivalents, will be a major goal. In relation to these fundamental control and robot learning questions, the further understanding of orbital dynamics in nonlinear systems ultimately aims at understanding how smartly designed system dynamics inherently program desired phenomenological motion and energy structures that can then serve as base systems to robot synthesis, learning and adaptation through interaction with the real world.
The third stream covers robot design. Inspired by the human sense of touch with its Golgi tendon - an organ for measuring muscle tension in the human musculoskeletal system - or tactile perception in the skin, designing robots with proprioceptive force/torque sensing in their joints and link structure, as well as sensory skins at the surface level, endows them with the tactile capabilities to sensitively and safely explore the environment and interact with it via tactile control. The research we conduct in this area considers the synergistic integration of morphology and design, sensory materials and systems, electronics, and control in synthesizing tactile robots. This quest ranges from developing torque-controlled drives and viscoelastic actuators to production assistants, surgical robots, tactile endoprostheses, and hands equipped with proprioceptive and tactile sensing, high-performance communication, and highfidelity interaction learning and control. Future systems should not only be equipped with Golgi-tendonlike sensitivity (proprioception) but also with the sense of touch (exteroception) provided, for instance, by novel artificial skins. The related sensing contingencies and the learning of related motor commands have to be realized in a novel bio-inspired “robot cerebellum”. Such a system will, for the first time, be able to technically replicate the unique physical exploration and interaction capabilities of humans and, therefore, provide the fundamental technology for countless new applications in robotics and beyond.
The Human
Physical human-robot interaction (pHRI) aims at enabling humans and robots to share their workspace and to get into direct contact. To realize this, ensuring human safety is the most essential requirement. In this context, our work on injury protection and safety in robotics unfolded into an entire new research discipline tightly connecting control and machine learning to injury biomechanics and accident research. My group developed the fundamental analysis and quantification of potential human injuries in pHRI, seeking to understand the inherent risks emanating from robot-human collisions and embedding these into the robot intelligence. Future work focuses on developing a general and commonly accessible atlas of human injury protection in robotics as a standard reference book and database. More injury protection data shall be collected, classified, and categorized to then build the ultimate algorithmic goal: the injury protection memory as the basis for a universal robot’s Sense of Safety.
Understanding human injury mechanisms helps develop injury protection methodologies in robotics; a more profound understanding of the “human neuromachine” can in turn feed back to advance robot design, control, and learning. The co-evolved human morphological building plan and motor control system are tightly coupled with each other and give humans their incredible abilities to learn how to grasp and manipulate the world. To understand the computational principles behind human motor learning, one needs to ask which representations do so elegantly link the cognitive to the motor control level produced by the central nervous system. The fundamental goal for our future research is to help further decode the relation between planning and control representation in the brain and its relation to synergy-based motor control and learning in human peripheral neurodynamics of movement and manipulation. For this, a unified energy-consistent human model shall be developed leveraging robotic modeling and identification algorithms that can be used as an in-silico testing system of motor control and learning representation hypotheses. For this, extensive fundamental brain-machine interfacing studies are necessary to laying also the foundation for brain-controlled symbiotic robot assistants.
The Discovery
A core research theme in our machine intelligence efforts is the development of novel machine learning paradigms that enable complex dynamical systems to learn its self, how to interact with the world, and how to autonomously acquire and generalize knowledge. For this, integrating nonlinear control theory and machine learning with intelligent data representations opens up entirely new perspectives and avenues in generating sophisticated machine intelligence. Although this discipline is still in its early days, there is a surprisingly high number of robotic problems that can be processed with existing results. One of our current core research topics is to develop collective learning algorithms to accelerate skill and knowledge learning using large-scale networked AI robot systems. This will involve integrating autonomous skill learning with learning-from-demonstration (LfD), reinforcement learning, and genetic algorithms with black-box optimization. In our framework, the collective comprises individual learning agents seeking a solution to their individual problems while concurrently sharing the found solutions with the other agents ultimately achieving exponentially fast learning. Regarding the developed skill learning framework, we plan to extend it to encompass a wide variety of different tasks, skill classes, and skill instances, such as opening doors, using tools, and other manipulation problems from human-centered environments and industries. Currently, we bring these approaches into the real world and enable robots to learn complex assembly tasks for state-of-the-art industrial production and manufacturing. Finally, we intensify our initial work on understanding how optimal control and machine learning can be unified under a single mathematical framework to learn and generalize optimal controls depending on sensory and contextual feedback. Ultimately, all these machine learning efforts funnel into tackling a decadal research problem, the generation of the Tree of Skills—the dual to the Tree of Robots—linking collective AI and tactile skill learning to a structured curriculum for machines to acquire, categorize, and synthesize real-world physical skills for health, manufacturing, or space.
The second decadal problem we focus on is autonomous scientific discovery, where we translate and extend our results in robot assistants, machine learning, and control into the experimental discovery processes of the natural sciences. In chemistry, medicine, biophysics, and biology, generating experimental knowledge still requires mostly time-consuming manual experiments in the laboratory (e.g., molecule synthesis, drug discovery, microscopic cell analysis, or individualized mRNA-based drugs in cancer therapy). The central scientific question we work on deals with cross-scale analysis and synthesis processes using intelligent robotic laboratory assistants that independently plan, modify, and evaluate experiments and massively accelerate their execution. The goal is for these self-learning multiscale robotic systems to develop maximum efficiency in conjunction with each other and in cooperation with humans. This will lead to a revolution in experimental evolving knowledge generation and systems synthesis in general. It will be possible to systematically answer questions such as how to control cell differentiation and fate (e.g., in cancer control), uncover the electromechanical interaction language of cells, or discover nanomaterials with optimized or possibly novel properties. We have already spent a decade developing and soon to be published next-generation AI-driven measurement and discovery systems—autonomous holotomography robots, electro-haptic telerobots, and cooperative discovery algorithms—key elements in algorithmizing the scientific method in the biomedical field.
Teaching Philosophy
I consider teaching to be the process of instilling fundamental concepts and paradigms for life-long learning. Once students leave the formal educational system they need to have the right skill set to handle a variety of challenges awaiting them during their careers, from the initial innovative idea or challenging real-world problem to the final solution. To achieve this goal, the first obligation of an academic instructor is to create a safe, interactive and collaborative environment that encourages students to be active members in the learning process. Equally important is to display one’s own sincere interest and passion toward the taught materials, stimulating students to not only display engagement but also avid enthusiasm for what they are learning. In fact, becoming such an inspirational teacher that supports and guides students to develop towards highly trained, curious and responsible engineers, scientists and thinkers has been the main reason that I chose to be an academic. My enthusiasm and passion for my profession also helped me to motivate my students to strive for excellence and develop the right sense of questioning the existing.
The two core pillars to my approach to train students are:
- foundational multi-disciplinary skills and
- the ability to find solutions to complex engineering problems.
Ultimately, I want to empower students to be able to systematically find the way from an initial innovative idea or challenging real-world problem to the final solution. Robotics and machine intelligence at large bridge diverse fields such as design, control, machine learning, sensing and artificial intelligence. While I believe, it is crucial to have a very strong background in one of these domains, the basic understanding and embracing of the full dimension and richness of our discipline is equally essential to achieve true innovation and unveil the unforeseen.
My courses thus build strong bridges between theoretical rigor and highly skilled implementation in complex robotic systems, ranging from mechatronics design and controller synthesis to learning algorithms and system integration. On a general note, robotics has transformed from being a specialized subfield of CS, EE and ME to an interdisciplinary field that soon will be considered an independent scientific discipline. Therefore, I believe that every engineering school should establishdedicated Robotics Master’s and PhD programs; in the long run also at the undergraduate level.