BiCAT: Improved Bimanual Coordination via Action Tokenization
November 16, 2026
BiCAT is a bimanual coordination action tokenizer that shares a single motion representation across both arms while conditioning on individual roles. This prevents the entanglement of cross-arm dependencies found in standard concatenated action vectors used by VLA models.
HOW THIS AFFECTS YOU
●
builderYou can achieve better coordination in dual-arm robots by using a shared motion repertoire.
●
researcherThis offers a cleaner way to encode multi-arm coordination in autoregressive models.