Generalizable Embodied Foundation Model

In this talk, we will start from the fundamental question of what is, how to measure, how to improve the generalization ability of deep neural networks, especially on the out-of-distribution generalization ability, where the test data distribution differs from the training data distribution. Then, we will demonstrate how to leverage the lessons from theoretical and algorithmic research for important applications, such as high dynamic range imaging, autonomous driving, control, and embodied artificial intelligence.