Talk by Prof. SiQi Zhou on the 8th of June
Speaker: Prof. Dr. SiQi Zhou
Date and Time: June 8th, 16:00-18:00, Vorhoelzer Forum
Title: Bridging Machine Learning and Control for Safe and Intuitive Human–Robot Collaborative Tasks
Abstract: As robots become an integral part of our daily lives, they need to continuously learn and adapt to operate safely in unstructured, human-centric environments. In the literature, well-established control techniques provide the foundation for designing high-performance robot systems with desired theoretical guarantees. However, their reliance on accurate dynamics models and well-characterized environments can lead to suboptimal performance or unsafe actions when facing real-world uncertainties. This challenge motivates the integration of machine learning into the traditional robot decision-making stack. Our approach is to leverage the expressiveness and reasoning capabilities of learned models to enhance the robot system while utilizing control-theoretic principles to ensure safe real-world deployment. In this talk, I will illustrate our approach in three interconnected directions: (1) safe learning-based control that leverages learning methods such as neural networks in combination with control theory to compensate for dynamic uncertainties and enable agile robot performance, (2) multi-agent coordination that utilizes distributed control frameworks as a safety filter for reliably deploying high-level task plans generated by large-scale learning models, and (3) perception-based safe decision-making that tightly couples metric-semantic understanding of the environment with control, allowing robots to reason contextually and act safely with “common sense.” These approaches are demonstrated through real-world experiments on various robot platforms, including quadrotors, manipulators, and mobile manipulators. I will conclude this talk with a summary of our broader survey and benchmarking efforts to advance safe real-world robot autonomy, along with an outlook on future directions rooted in principled integration of learning and control for designing robot systems capable of making safe and context-aware decisions in human-centric environments.