Reliable intelligence requires feedback at all levels.
We advance the scientific foundations of intelligent systems that safely interact with humans and the physical world.
Physical AI systems, such as robots or autonomous vehicles, are structured hierarchically: from low-level control and learned policies, through perception and planning based on transformers and other learned models, to reasoning. Across these levels, the alphabet of feedback is extended: from continuous measurements to images and video, semantic information, language, and human intent. Whether and how the principles of feedback, prediction, optimization, and guarantees extend across this expanding alphabet is the central scientific question driving our research.
Physical AI: feedback across levels of abstraction
Future intelligent systems will not merely process information. They will continuously interact with humans and changing physical environments: robots move through and manipulate their surroundings, and autonomous vehicles and drones operate amid constant change. These systems must perceive, predict, reason, decide, act, learn, and adapt under uncertainty while remaining safe and reliable.
Physical AI closes feedback loops in the physical world through an ever richer alphabet of signals. Classical control primarily operates on continuous variables; Physical AI extends feedback to increasingly abstract representations, including probabilistic beliefs, images and video, semantic representations, symbolic knowledge, language, preferences, and human intent.
Feedback becomes hierarchical.
Future Physical AI consists of hierarchical intelligent systems connected through multi-level feedback loops. We study hierarchical feedback architectures spanning physical interaction, control, learning, planning, reasoning, and human decision making, with loops that cut across levels rather than following a strict cascade.
The question is open.
Whether the principles that made feedback reliable (stability, robustness, and performance guarantees) extend across these increasingly abstract representations is an open scientific question. Answering it defines our research program.
Our contribution.
Our work rests on a strong foundation in systems and control theory, built over decades of research in feedback control and Model Predictive Control. On this basis, we develop mathematical foundations for reliable Physical AI: systems that learn, reason, and act in the physical world while remaining safe, adaptive, and trustworthy.
Scientific breakthroughs in Safe Physical AI will not emerge from a single discipline but from the integration of systems theory, control, robotics, machine learning, communication, and engineering. TU Darmstadt provides one of Europe's strongest environments for this integration.
From Model Predictive Control to Physical AI
Model Predictive Control demonstrated that prediction, optimization, feedback, and constraint satisfaction can be combined within a flexible, mathematically rigorous closed-loop architecture that provides guarantees.
Our research extends this paradigm beyond continuous dynamical systems towards hierarchical feedback architectures operating on learned policies, world models, semantic reasoning, and human interaction. These levels do not form a strict cascade: information from the physical world reaches learning, planning, and reasoning directly, while higher-level decisions continuously reshape lower-level behaviour. Modularity is how such architectures master complexity: properties are established for individual components and composed across their interfaces and interconnections.
Grounded in results.
Our work builds on deep expertise in systems and control theory, predictive control, and constraint satisfaction, demonstrated through certified learning-enabled control, Bayesian optimization and surrogate-based decision making, uncertainty-aware prediction and control, embedded real-time optimization, and the industrial deployment of these methods with partners including Volkswagen, Bosch, Airbus, and Baker Hughes.
Core scientific pillars
|
Feedback & Control Feedback is the organizing principle of intelligent physical systems. systems & control theory · dynamic systems · learning-based control · hierarchical & networked control · state and parameter estimation · embedded real-time control |
|
Learning & Representations Learning extends feedback rather than replacing it. physics-informed machine learning · learned policies · multimodal learning · predictive models · generative diffusion & flow models · vision-language models (VLMs) · vision-language-action models (VLAs) · foundation models |
|
Optimization & Decision Making Principled decisions under constraints and uncertainty, from milliseconds to mission horizons. model predictive control · optimal control · planning · Bayesian optimization · policy optimization · robust optimization · decision making under uncertainty |
|
Guarantees and Safety Across Levels despite Uncertainty Our signature: guarantees that hold across learning, adaptation, and deployment. uncertainty quantification & propagation · robustness · compositional guarantees · reachability · verification · runtime assurance · certified learning |
Human-centered by design
Humans participate in feedback loops at every level. They define goals and constraints, provide demonstrations and preferences, supervise decisions, and interact physically with the systems we study. We therefore treat human-in-the-loop learning, shared autonomy, assistive and wearable robotics, and transparent decision support as integral components of Human-Centered Safe Physical AI, not as an interface added afterwards.
Physical application platforms
Our methods are developed for and validated on real physical systems and the digital infrastructure connecting them, spanning robotics and autonomous systems, energy, chemical and industrial processes, mobility, healthcare, communication, and large-scale scientific facilities.
-
![]()
-
Humanoid & Mobile Robotics
Locomotion, manipulation, exploration, and safe learning. -
![]()
-
Exoskeletons & Wearables
Assistance, shared autonomy, human co-adaptation. -
![]()
-
Aerial Systems
Agile flight, decentralized control, planning, and estimation under uncertainty.
-
![]()
-
Autonomous Vehicles
Motion planning, personalization, vision-language-action models, and flow models.
-
![]()
-
Communication & Networked Systems
Resilient feedback over networks, control and communication co-design, and cyber security.
-
![]()
-
Distributed Edge Intelligence & Embedded Control
Learning, control, and virtual sensors across embedded devices.
-
![]()
-
Healthcare & Medical Engineering
Physiological systems, brain-machine interfaces, closed-loop deep brain stimulation, and personalized therapies.
-
![]()
-
Batteries, Fuel Cells & Energy Systems
Modeling, state-of-charge and state-of-health estimation, fast charging, and resilient energy networks.
-
![]()
-
Chemical & Biochemical Processes
Optimal operation and monitoring of chemical and biotechnological processes.
-
![]()
-
Industrial Automation & Processes
Automation, optimal operation, monitoring, and virtual sensors for production and process systems, including autonomous drilling.
-
![]()
-
Molecular Manipulation:
Scanning probe platforms and feedback control at the nanoscale.
-
![]()
-
Control & Optimization of Accelerators
Optimization, machine learning, automation, and feedback for large-scale instruments.
