MA-VLA: Multi-Arm Vision-Language-Action Model for Collaboration and Compositional Generalization
Zaibin Zhang, Junlan Xiao, Zhongbo Zhang, Yifan Wang, Li Kang, Yiran Qin, Changxing Xia, Heng Zhou, Talas Fu, Enshen Zhou, Ruimao Zhang, Zhenfei Yin, Huchuan Lu, Lijun Wang
ECCV 2026
cs.RO
2026-08-26
MA-VLA decomposes instructions into per-arm atomic actions and shuffles arm roles; unseen collaborations go from 0% for Pi0 to 13% average success.
Multi-arm collaboration is becoming a default setting in embodied manipulation, yet most VLAs still treat language as one global command. Who does what is inferred from role statistics in the data. If the left arm always hands and the right arm always receives in training, swapping roles or changing the stacking order at test time collapses the policy. Groups at Dalian University of Technology and Oxford tighten the question: the atomic actions are in the training set, the collaboration structure is not. Can the model recombine them? They call this multi-arm compositional generalization.
MA-VLA has two stages and stays end-to-end at inference.
The planner is GPT-4.1, not fine-tuned. Given a high-level instruction and current images, it picks stage-wise, per-arm mid-level prompts from a finite vocabulary (grasp, hold bowl, place cube) and canonicalizes the output so every prompt lands in that set.
The executor starts from official pi0base and emits continuous actions with flow matching. Atomic prompts for all arms are concatenated into one string, "Arm0: …, Arm1: …", then consumed with global views, wrist views, and proprioception in a single forward pass. A multi-head projection splits the shared latent into per-arm actions.
Two training-time perturbations leave the behavioral cloning objective unchanged. Arm Shuffle permutes, with probability pshuffle, each arm's state, view, prompt, and action binding, so the model has to read the prompt instead of memorizing arm indices. View Dropout zeroes a subset of cameras with probability pdrop. Each simulation task has 150 expert demos; a rule-based parser labels frame-level atomic prompts from contact, grasp, and pose thresholds.
In-domain RoboFactory, 100 rollouts, mean success:
| Setting | Pi0 | MA-VLA |
| Two-arm four-task avg | 80.3% | 83.5% |
| Three/four-arm four-task avg | 76.5% | 83.3% |
Three-arm Stack Cube moves 48% to 58%, four-arm Long Pipeline Delivery 82% to 97%. On RoboTwin 2.0 Hard the average is 41.1 to 49.0, Handover 27 to 46, Cabinet 59 to 71. Stamp Seal stays low, 3 to 12.
Out-of-domain is the headline. Unseen cube-stacking orders, Pass Two Shoes, and Stack Two Bowls: DP, Pi0-FAST, and Pi0 all score 0. MA-VLA averages 13.0, with 28 on the GBR order and about 9 to 10 on the other three.
Real dual-arm SO101, 50 teleop demos per task, 20 evaluation episodes:
| Task | Pi0 ID/OOD | MA-VLA ID/OOD |
| Stack Bowls | 9/20 / 0/20 | 12/20 / 10/20 |
| Place Cubes | 12/20 / 0/20 | 15/20 / 8/20 |
| Pass Toys | 3/20 / 0/20 | 6/20 / 2/20 |
| Stack Cubes | 5/20 / 0/20 | 8/20 / 2/20 |
Ablation on unseen three-arm stacking order: atomic prompts alone take in-domain 48 to 58 and leave OOD at 0. Adding Shuffle yields 7.3 OOD and 52 in-domain. Adding View Dropout reaches 15.3 OOD and 53 in-domain. Training a separate Pi0 per arm hits 61 in-domain and still 0 OOD, with model count growing linearly.
The bottleneck in multi-arm VLAs is often not the action head. It is the binding between role and arm index. Writing the division of labor as recombinable mid-level language, then permuting identity at train time, is what moves OOD success off zero. In-domain gains are a few points. The story is compositional transfer.
Reproduction needs two extra pieces: an atomic-action vocabulary with rule-based labels, and a GPT-4.1 planning call. The executor is a Pi0 finetune, not a new architecture.
No standalone limitations section. OOD average is 13%, and real-robot Pass Toys / Stack Cubes OOD is 2/20. Recombination is not the same as a usable policy. The planner does not learn; it depends on GPT-4.1 and a hand-built vocabulary, so collaboration patterns outside that set cannot be planned. Shuffle rates are tuned per OOD task (0.6 / 0.8 / 1.0) with no universal hyperparameter. Simulation baselines adapt single-arm VLAs and retrain them, which is fair, but the comparison is against adapted Pi0, not a native multi-arm method. Twenty real episodes is a tight sample.