← Paper Digest

BiCAT: Bimanual Coordination Action Tokenization

Xuanwei Liu, Guipu Cao, Zhongyi Luo, Jiafei Duan

November 2026

Bimanual YAM capping a pen and packing a box, sped up 2.5x and 4x.

TL;DR BiCAT is an action tokenizer for bimanual VLA policies that gives each arm its own token stream, computed with access to the other arm. It raises mean success on ten RoboTwin 2.0 tasks by 17.6 points over the best of eight tokenizers and succeeds in 113 of 150 real-robot trials, against 74 for OAT.

What they do

Autoregressive VLAs turn action chunks into discrete tokens, and most tokenizers treat a bimanual robot’s two arms as one concatenated vector. BiCAT tokenizes each arm separately but lets each arm’s tokens see its partner. The policy then predicts the left arm’s tokens, followed by the right arm’s tokens conditioned on them.

How it works

  • Mirror canonicalization. A per-robot sign vector flips the right arm’s joints into the left arm’s convention, and both arms share normalization bounds, so mirrored motions map to the same trajectory.
  • Shared encoder with cross-arm attention. One causal temporal encoder compresses each arm’s 32-step chunk into 8 latents. Symmetric causal cross-attention adds partner context before finite scalar quantization with 1920 codes shared by both arms.
  • Joint decoder. A Transformer decoder rebuilds both trajectories from the paired codes. The tokenizer is trained on reconstruction alone, then frozen.
  • Two-block prediction. A VLM (PaliGemma 2-3B in simulation, Molmo2-ER on real robots) predicts the 8 left-arm tokens in parallel, then the 8 right-arm tokens, so each chunk takes two decoding steps.

Results

  • RoboTwin 2.0, ten tasks, 100 rollouts each: 69.3% mean success versus 51.7% for NAC, the best of eight baselines. The gain is 1.5 points on single-arm tasks, 17.5 on loosely coupled and 21.3 on tightly coupled tasks.
  • 40.5 ms per chunk on an H200, 5.6x to over 11x faster than tokenizers decoded token by token.
  • Ablations: without cross-arm attention (51.7%) or with both arms in one block (51.1%), BiCAT falls to NAC’s level; without the joint decoder it reaches 56.0%.
  • Real robots (bimanual SO-101, Unitree G1, bimanual YAM; six tasks, about 100 teleoperated demos and 25 trials each): 113/150 successes versus 74 for OAT and 25 for FAST, best on all six tasks, at 91.7 ms per chunk versus 434.1 ms for FAST.

Limitations

BiCAT works in joint space; its fixed per-joint sign mapping does not carry over to end-effector actions. Each tokenizer is trained for a single embodiment and gives both arms equal, fixed token budgets. In simulation it trails other tokenizers on Handover Mic (88% versus 94% for OAT) and Dump Bin Bigbin (85% versus 87% for NAC).

Videos

Unitree G1 moving a bottle and stacking a bowl on a plate, sped up 2.5x and 3x.

Figures

From Figure 1: Left: existing tokenizers treat the two arms as a single arm and mix both arms in every token; BiCAT g...
From Figure 1: Left: existing tokenizers treat the two arms as a single arm and mix both arms in every token; BiCAT gives each arm its own token stream, aware of its partner. Right: BiCAT rollouts on three embodiments with each arm's motion over time, and success rates in simulation and on real robots; white numbers give the gain over the best baseline.
From Figure 2: Overview of BiCAT. Top: mirror canonicalization flips the signed joints of the right arm and applies s...
From Figure 2: Overview of BiCAT. Top: mirror canonicalization flips the signed joints of the right arm and applies shared normalization; the shared encoder and causal cross-attention give each arm its partner's context before FSQ, and the joint decoder reconstructs the chunk under a reconstruction loss. Bottom: a VLM predicts the left-arm tokens, then the right-arm tokens; the frozen decoder returns actions, executed on Bimanual SO-101, Unitree G1 and Bimanual YAM. Teal: left arm; orange: right arm.
From Figure 4: BiCAT rollouts on three real bimanual embodiments. Each row shows four keyframes of one task with a sh...
From Figure 4: BiCAT rollouts on three real bimanual embodiments. Each row shows four keyframes of one task with a short task description.
Abstract

Autoregressive vision-language-action (VLA) models predict discrete action tokens with the next-token objective of their vision-language backbone, so the action tokenizer defines the interface between the policy and the robot. In bimanual manipulation, the two arms share a repertoire of movements yet coordinate through complementary roles, and an action representation should capture both. However, most existing tokenizers encode both arms as a single concatenated action vector, in which motion patterns common to the two arms are not shared and cross-arm dependencies remain entangled within each token. The policy must therefore learn both aspects implicitly. We propose Bimanual Coordination Action Tokenization (BiCAT), which shares one motion representation across arms while conditioning each arm's tokens on its partner through symmetric cross-arm attention and a joint decoder. The policy then predicts all tokens of each arm in parallel, generating the left arm first and the right arm conditioned on it. Across ten RoboTwin 2.0 tasks and six real-world tasks on three embodiments, BiCAT achieves the highest success among the compared tokenizers. In simulation, it improves mean success by 17.6 percentage points over the strongest of eight baselines, with the largest gains on tasks that require close coordination between the arms, while requiring only 40.5 ms per action chunk. These results suggest that the action tokenizer is a natural place to model bimanual coordination.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.