In this informal talk, I will share a few small tricks developed in my past research on trustworthy autonomy. These span applications from autonomous driving to intelligent defect inspection, and include approaches such as black-box testing, randomized smoothing for robustness, safety-aware optimization in object detection, formal verification, and out-of-distribution detection. Building on these experiences, I then turn to the emerging domain of service robotics, where generative AI such as LLMs and VLMs increasingly shape the robot’s cognitive core. The talk alternates between concrete methods I have devised for autonomy based on task-specific neural networks (e.g., object detectors) and the new challenges that arise when transferring them to autonomy powered by generative models. The goal is to reflect on what carries over, what breaks down, and how we can chart new directions toward trustworthy autonomy in the era of generative AI.

Chih-Hong Cheng is a professor at the University of Oldenburg, Germany. Previously, he was a tenured faculty member at Chalmers University of Technology, Sweden, where he continues to hold a part-time appointment. His research interests include software engineering, formal methods, and AI/ML for trustworthy autonomy. He received his doctoral degree in computer science from the Technical University of Munich. After his PhD, he worked primarily in government (fortiss, Fraunhofer IKS) and industrial research centers (ABB, DENSO), while also holding interim professorships at TU Munich and the University of Hildesheim.