Online and Hybrid System Identification Directly From Raw Sensory Signals

Explore cutting-edge techniques for system identification in adaptive engineering and experimental sciences through this insightful lecture. Delve into the challenges of extracting structured models from sparse observations and rich sensory data, particularly in robotics and autonomous systems. Discover how neural networks with physical model inductive biases are revolutionizing this field. Learn about switching density networks for learning hybrid controllers in robotics, and examine the application of variational recurrent neural network architectures for efficient parameter estimation from video streams. Investigate the versatility of these methods as they are applied to complex domains like molecular geometry modeling, including the inversion of ultrafast X-ray scattering with dynamics constraints. Gain valuable insights into the latest advancements in online and hybrid system identification, directly applicable to various fields of engineering and scientific research.