Ordered Action Tokenization Improves Robot Policy Learning
A team of researchers has introduced Ordered Action Tokenization (OAT), an innovative technique designed to convert continuous sequences of robot actions into structured, ordered tokens. Current methods either generate lengthy token sequences through analytical means or fail to provide a structured approach in learned tokenizers. OAT meets three key criteria: efficient compression, complete decodability, and a well-organized token space. It employs a transformer that incorporates registers, finite scalar quantization, and training that promotes order, allowing initial tokens to capture broad control data while subsequent tokens refine finer details. This organized tokenization aims to enhance compatibility with downstream visuomotor policies.
Key facts
- OAT is a learned action tokenizer for robot action chunks.
- It satisfies high compression, total decodability, and ordered token space.
- Uses transformer with registers, finite scalar quantization, and ordering-inducing training.
- Early tokens encode coarse control, later tokens refine residual details.
- Addresses limitations of analytical and latent tokenization methods.
- Aims to improve visuomotor policy learning.
- Published on arXiv as 2607.21670.
- Action tokenization maps continuous robot actions to discrete tokens.
Entities
—