Researchers led by Xuanwei Liu have built BiCAT, a way of turning the movements of a two-armed robot into the discrete tokens that vision-language-action models predict, according to a paper on arXiv. Most tokenizers treat both arms as one long action vector, so the policy has to learn on its own which movements the arms share and how they depend on each other.
BiCAT uses one shared motion representation for both arms and links each arm's tokens to its partner's through attention between the arms and a joint decoder. The policy predicts the left arm's tokens first and then the right arm's, each arm's tokens in parallel.
Across ten RoboTwin 2.0 simulation tasks and six real-world tasks on three robots, BiCAT had the highest success of the tokenizers compared. In simulation its mean success was 17.6 percentage points higher than that of the best of eight other tokenizers, with the largest gains where the arms must work closely together, and it needs 40.5 milliseconds per chunk of actions.



