NUS's BiCAT action tokenizer for bimanual robots: 113/150 real-robot wins, 2.7x lower latency than OAT
DJiafei · x · 2026-11-18
Researchers at the National University of Singapore propose BiCAT (Bimanual Coordination Action Tokenization), a new action tokenizer for bimanual VLA models: instead of concatenating both arms into one vector, each arm gets its own token stream with a shared motion vocabulary.
- Method: mirror canonicalization flips the right arm's signed joints so mirrored motions share tokens; symmetric cross-arm attention conditions each arm on its partner before quantization; a joint decoder rebuilds both arms in two steps (40.5 ms per action chunk on H200).
- Simulation: 69.3% mean success across 10 RoboTwin 2.0 tasks, +17.6 pts over the best of 8 tokenizers (OAT, NAC, FAST, QueST, etc.); +21.3 pts on tightly coupled tasks.
- Real robots: 113/150 trials succeeded across 6 tasks and 3 embodiments vs 74 (OAT) and 25 (FAST); 4.7x faster than FAST, 2.7x lower latency than OAT on YAM (91.7 ms vs 251.6 ms).
Paper, code, and project page are public.
Related event: NUS Introduces BiCAT, a Bimanual Coordination Action Tokenizer for Robots(4 posts)→
More from Embodied
- Rodney Brooks, Roomba creator, calls today's humanoids learning dexterity 'pure fantasy' — lukas_m_ziegler · 2026-11-18
- Two arms, two token streams, one shared vocabulary: new robot manipulation work open-sourced — 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
- San Bernardino County deploys autonomous firefighting drones to kill wildfires early — import_jmr · 2026-11-18
- Vibe-coded AirDrop turns a photo into a walk-in 3D scene on Vision Pro — Scobleizer · 2026-11-18