When: October 23rd 2025, 1:00pm AEDT
Where: This seminar will be partially presented at the ACFR seminar area, J04 lvl 2 (Rose St Building) and partially online via Zoom. RSVP
Abstract:
Learning-based controllers have demonstrated remarkable empirical performance in robotics, yet they remain fundamentally limited by their lack of transparency and formal stability guarantees. This gap has motivated extensive research into verification methods, but existing approaches either impose significant training constraints or demand prohibitive computational resources. We address this foundational challenge by developing a new class of neural dynamical systems that embed stability guarantees directly into their structure, eliminating the need for expensive post hoc certification. Our framework enables the design of controllers that are provably exponentially stable or passive by construction, thus bridging the critical divide between the flexibility of learning-based methods and the formal rigor required for safety-critical applications. We demonstrate that this approach maintains the representational power necessary for complex robotic tasks while providing verifiable stability properties, and we establish theoretical conditions that formally guarantee these safety attributes. This work opens new directions for integrating principled stability constraints into neural control design, potentially transforming how we develop trustworthy learned controllers for real-world robotics.
Brief Bio:
Dechuan Liu is a PhD student in the ACFR Control Group at the University of Sydney, supervised by Professor Ian Manchester. He completed a Bachelor of Engineering and a Bachelor of Science at the University of Sydney in 2024. Dechuan’s research focuses on learning-based control for robotics, integrating machine learning with control theory to develop methods that offer theoretical guarantees such as stability and robustness.