ELLIS header
University of Stuttgart Logo
Max Planck Institute for Intelligent Systems Logo

Distinguished Lecture Series - Talk by Georgia Chalvatzaki (TU Darmstadt)

We are pleased to announce our upcoming Distinguished Lecture Series talk by Georgia Chalvatzaki (TU Darmstadt)! The talk will take place in person on December 2nd, from 9:45 - 11:15 CET, in room UN32.101., Universitätsstraße 32, Campus Vaihingen, University of Stuttgart

Georgia Chalvatzaki is a Full Professor of Interactive Robot Perception and Learning at TU Darmstadt, Germany, where she leads the PEARL Lab. Her research focuses on robot learning, combining structured representations, geometric and physical inductive biases, and learning from interaction to enable robots to acquire generalizable manipulation skills. She is a recipient of an ERC Starting Grant (2024), the Alfried Krupp Sponsorship Prize (2025), and a DFG Emmy Noether Grant (2021). She currently serves as Program Chair of CoRL 2026 and Vice President of the German Robotics Society (DGR).

Title: The Structure of Physical Interaction: Towards Reliable Robotic Embodied Intelligence

The Structure of Physical Interaction: Towards Reliable Robotic Embodied Intelligence

Large pretrained models and growing robot datasets have substantially broadened what robots can do, yet robust sensorimotor competence in contact-rich, partially observable, and novel situations remains elusive. I will argue that exploiting the structure of physical interaction is complementary to scale, and present structured robot learning as a research program built on this view. By structure I mean more than an architectural prior: structure shapes what a robot senses, how it represents the world, how its training data is generated, and how learning itself is organized. Appropriate structure is particularly valuable where interaction is limited, costly, or safety-critical, because it can supply what data alone recovers poorly: invariances, physical consistency, representations of uncertainty, and the compositional organization of tasks. It can also change how effectively additional data translates into generalization. Yet structure is not uniformly beneficial, and I will discuss when it helps and when an inaccurate prior harms learning. I will organize the talk around the kinds of structure that embodied agents need. Geometric and morphological symmetries enable sample-efficient and transferable policies, with benefits that depend on the robot’s embodiment. Physical structure, through tactile and multisensory learning, captures interactions that vision alone cannot reliably resolve. Informational structure concerns what a representation should preserve: explicit representations of uncertainty in pose and scene support perception and decision-making under occlusion and ambiguity, while features distilled from pretrained diffusion models retain fine-grained spatial information for contact-rich manipulation that common discriminative pretraining objectives tend to suppress. At the task level, semantic-geometric representations enable compositional planning from human demonstrations, and learned interaction dynamics support physical human-robot collaboration. Physical interaction constrains not only what robots should represent, but also how their learning should be organized. I will show how structure in data generation and in the learning process enables scale and increasingly autonomous learning. Generalizable reward design allows reinforcement learning to generate synthetic training data for language-conditioned bimanual dexterous manipulation, improving policy performance as the synthetic dataset grows and enabling zero-shot transfer to real robots. Structured workflows with explicit verification allow general-purpose AI agents to orchestrate robot reinforcement learning, while autonomous real-world practice allows robots to improve their skills through physical interaction. I will close with a question that I consider central for the next stage of robot learning: which structures should be built into learning systems, which can be extracted from large-scale data, and which must robots acquire through their own physical experience? I will argue that answering it requires a shift from structure designed by researchers to structure that robots discover, test, and refine through interaction, while remaining safe as they learn. To conclude, I will outline how structured world models, causal representations, and mechanisms for reliable adaptation could make this shift possible.

Date: December 2nd, 2026
Time: 9:45
Place: Universitätstraße 32, Room 101, Campus Vaihingen of the University of Stuttgart.

Looking forward to seeing you all there! No registration necessary.