Tag
This paper introduces ActionPiece, a novel action tokenization method for autoregressive vision-language-action models that uses physical rank consistency to improve the fidelity of action relationships, evaluated on benchmarks like LIBERO and SimplerEnv.