Stable Robot Motions on Manifolds: Learning Lyapunov-Constrained Neural Manifold ODEs.

Neural ordinary differential equations provide a continuous-time, expressive framework for learning robot dynamics from data, making them well suited for motion generation. Recently, neural ODEs have been extended to Riemannian manifolds, the geometric structures that many important robotic quantities such as rotations and stiffness lie on. In this talk, I will presents a framework that combines neural ODEs on Riemannian manifolds with learned Lyapunov functions to provide stability guarantees. Stability is essential for ensuring robust task completion in motion generation.

David Boetius is a PhD student at University of Konstant working on formal methods for neural networks under the supervision of Stefan Leue and Tobias Sutter. Recently, he has developed an an interest in cyber-physical systems.