[rl-list] Safe RL Talk Invitation: Integrating Machine Learning and Control for Safe and Efficient Robot Decision-Making

[Josip Josifovski]

Dear colleagues,

On Thursday (3rd of April 2025) we are hosting a talk in the scope of the Safe Reinforcement Learning Online Seminar. The details for the talk are given below. We invite you to join via zoom if you are interested and feel free to share this invitation with colleagues or students that might find the talk relevant.

Talk Title: Integrating Machine Learning and Control for Safe and Efficient Robot Decision-Making
Speaker: Dr. SiQi Zhou (Technical University of Munich)

Host: Dr. Shangding Gu
Talk Time: 03 April 2025, 18:00h CEST Time (Amsterdam, Berlin, Rome, Stockholm, Vienna) / 09:00h California time / 12:00h Eastern Time / +1 00:00h Beijing time. Download a calendar event for easy import at this link.

Join Zoom Meeting:
https://tum-conf.zoom-x.de/j/67269576319?pwd=ubowFTwtJCpqeW7EYCb7fbfKpFgthd.1  
Meeting ID: 672 6957 6319
Passcode: 903926

Abstract: Robots are envisioned to become reliable human companions in domains ranging from industrial applications to our daily lives. In the literature, well-established control techniques provide the foundation for designing high-performance autonomous robot systems with desired theoretical guarantees. However, these control techniques often rely on a dynamics model of the robot, and any inaccuracies of the model can result in suboptimal performance or even unsafe actions. This limitation motivates incorporating learning into the traditional robot decision-making software stack. In our work, departing from control theory, we develop neural control approaches that safely and efficiently exploit the expressiveness of neural networks to enhance the performance of robots in uncertain environments. This talk will encompass a set of our neural control work ranging from offline inverse dynamics learning for improving the performance of robots to online Lipschitz network adaptation for closing the model-reality gap in uncertain robot systems. We demonstrate our approach in real-time robot experiments, including quadrotor impromptu trajectory tracking and flying an inverted pendulum. In this talk, I will also introduce our review paper on safe learning in robotics and related benchmarking efforts that are targeted to provide an overarching view of the recent advances in learning-based control and reinforcement learning for robotic applications. I will conclude by sharing our latest work that closes the perception-action loop with a semantic understanding and thereby enables robots to safely operate in everyday scenarios.

Bio: SiQi Zhou is a Senior Scientist at the Chair of Safety, Performance and Reliability for Learning Systems at the Technical University of Munich (TUM). She received her Ph.D. from the University of Toronto in 2022 and her B.A.Sc. degree from the University of Toronto Engineering Science program in 2016. Her research lies at the intersection of robotics, machine learning, and systems control. By integrating learning techniques and control theory, she aims to develop approaches that safely and efficiently improve the performance of robots in uncertain and unstructured environments. SiQi was selected as one of the MIT Rising Stars in Aerospace (2021), an RSS Pioneer (2022), and a recipient of the EU Marie Skłodowska-Curie Actions (MSCA) Fellowship (2024).
 
Safe RL Online Seminar Homepage: https://sites.google.com/view/saferl-seminar/home, We welcome the researchers and students who are interested in safe RL to join us! To receive relevant seminar information in time, please click the link to register.

Best regards,
Josip on behalf of the Safe Reinforcement Learning Online Seminar organizers

Josip Josifovski

Technical University of Munich (TUM)
School of Computation, Information and Technology
Chair of Robotics, Artificial Intelligence and Real-time Systems (I6)
Schleißheimer Str. 90A, 85748
Garching bei München
https://www.ce.cit.tum.de/air/people/josip-josifovski-msc/