2026Harvard University and Stanford Universityopen source
OAT: Ordered Action Tokenization
v1.0
OAT is a learned action tokenizer designed to enhance autoregressive policies in robot learning by discretizing continuous action sequences into ordered tokens. It addresses challenges in action tokenization by providing reasonable compression, universal decodability, and a left-to-right causally ordered token space.
Transformer-based register tokens
Finite scalar quantization
Ordering-inducing training mechanisms