Researchers introduce ActionPiece, a method for preserving physical action relationships in autoregressive vision-language-action models through joint supervision of representation learning and quantization. This approach improves action tokenization and policy learning, achieving high accuracy on various benchmarks.