Action Chunking with Transformers, piece by piece
ACT is the usual starting point for modern imitation learning. This post walks through action chunking, the architecture, the style variable, training and inference, how attention works inside it, and temporal ensembling.