BiCAT gives each robot arm its own token stream instead of one flat vector
DJiafei · x · 2026-11-18
Part 2 of the BiCAT thread lays out the core design: today's tokenizers flatten both arms into one vector, while BiCAT gives each arm its own token stream.
Human arms don't move independently — they share a repertoire of movements while taking complementary roles (one holds, one folds; one passes, one receives). BiCAT brings this structure into a VLA's action tokens, sharing motion patterns across arms while explicitly modeling their coordination.
More from Embodied
- NUS's BiCAT action tokenizer for bimanual robots: 113/150 real-robot wins, 2.7x lower latency than OAT — DJiafei · 2026-11-18
- Two arms, two token streams, one shared vocabulary: new robot manipulation work open-sourced — DJiafei · 2026-11-18
- BiCAT mirrors the right arm's signed joints so both arms share motion tokens — DJiafei · 2026-11-18
- BiCAT builds bimanual coordination directly into a VLA's action tokens — DJiafei · 2026-11-18
- Startups put Meta glasses on day laborers to harvest first-person video for robotics companies — hiclemi · 2026-11-18
- NVIDIA RTX Spark compact desktops ship with 128GB unified memory for local AI agents — NVIDIA Blog · 2026-11-18