Modern Robot Learning from Scratch - Lecture 3 - Deep Generative Modeling | Variational AutoEncoders | Modern Robot Learning from Scratch

In this lecture, we look at deep generative modeling. This is an excellent alternative for the multi-modality problem of point estimate policies in Imitation Learning, which we looked at in the last lecture. We will look at a specific deep generative model called Variational Autoencoders (VAEs) which forms the heart of the Action Chunking with Transformers (ACT) policy. We will develop a very strong intuition about VAEs and learn about them from scratch, including a practical example. This lecture or the theoretical will form a very solid foundation for the Action Chunking with Transformers policy understanding.