Title: Physics-informed machine learning: Opportunities and Challenges for Robotics
Speaker: Prof. Subramanian Ramamoorthy
Abstract: Models have held an important but controversial position with the AI community, with many valid arguments for and against their use. In applications that require high-performance autonomous system behaviours and assurance of safety, model-based approaches still hold the dominant position, although it is now becoming possible to blur the boundaries between model-based and data-driven or trial-and-error paradigms.
In this setting, there is a crucial role for system and model identification from sparse observations. Similar problems arise in the experimental sciences in the form of inverse problems. In this talk, I will present some recent results from a programme of work involving neural networks with inductive biases in the form of physical models.
I will start from the basics, by motivating the need for injecting structure into data-driven models - using as an example robotics problems of learning hybrid controller descriptions from demonstration. Next, I will describe how variational recurrent neural network architectures can be used to perform data-efficient parameter estimation with dynamics observed in video streams. I will show that these same methods can be applied in settings involving reasonably complex dynamical phenomena. So, I will describe how the same variational RNN architecture is able to address measurement inversion of molecular geometry from ultrafast X-ray scattering. Finally, I will end with a discussion of preliminary results from ongoing work on the use of simulation models to represent continuum phenomena with soft materials.
Bio: Prof Subramanian Ramamoorthy holds the Chair of Robot Learning and Autonomy in the School of Informatics, University of Edinburgh, where he is also the Director of the Institute of Perception, Action and Behaviour. He is an Executive Committee Member for the Edinburgh Centre for Robotics and Turing Fellow at the Alan Turing Institute. He received his PhD in Electrical and Computer Engineering from The University of Texas at Austin in 2007. He has been an elected Member of the Young Academy of Scotland at the Royal Society of Edinburgh, and Visiting Professor at Stanford University and the University of Rome “La Sapienza”.
His research focus is on robot learning and decision-making under uncertainty, with particular emphasis on achieving safe and robust autonomy in human-centred environments.
Between 2017 - 2020, he served as Vice President - Prediction and Planning at FiveAI, a UK-based startup company developing autonomous vehicles technology. He continues to be involved with the company as Scientific Advisor.