A new open-source tokenizer wants to fix how robots translate camera and language input into movement.
Researchers built M2Tok, a discrete action tokenizer for Vision-Language-Action (VLA) models, the systems that let a robot arm turn a camera feed and a text instruction into motor commands. Instead of using one shared codebook to compress raw motion into digital tokens, M2Tok splits the action signal into multiple heads, each implicitly tied to a different part of the movement, then gives each head its own codebook. That combinatorial setup lets the tokenizer represent finer motion detail with less loss. The team tested it on the RoboTwin and Simpler-Env simulation benchmarks plus three real-world tasks, and released the code on GitHub.
This targets a real weak point in robot AI: today's tokenizers often blur fine motor detail when they compress continuous movement into discrete tokens, capping how precise a downstream robot can be no matter how good the underlying model is. A tokenizer that reconstructs motion more faithfully should, in principle, raise that ceiling without retraining the language model itself.
It is a plumbing fix, not a new capability - and three real-world tasks is a thin sample for claims about closing a bottleneck. Whether multi-codebook tokenization holds up at the scale of a real warehouse or kitchen is still an open question.