Principle-centric Machine Learning for Embodied Foundation Models - Dr. Adrien Gaidon CVPR SDAS 2023
Can Machine Learning's success on the web translate to intelligent machines physically interacting with the real world, especially for safety-critical systems like cars and robots? In this talk, Adrien describes their unique approach to the problem of Embodied Intelligence: Principle-centric ML. It consists in going beyond data by complementing it with guiding principles to increase safety, generalization, sample efficiency, and robustness. Adrien showcases a few examples of our research leveraging physical and cognitive principles via self-supervised objectives, inductive priors, and simulation to not just let machines learn from examples but actually guide them as teachers toward better outcomes.