MA-VLA: Multi-Arm Vision-Language-Action Model for Collaboration and Compositional Generalization
Authors: Zaibin Zhang, Junlan Xiao, Zhongbo Zhang, Yifan Wang, Li Kang, Yiran Qin, Changxing Xia, Heng Zhou, +6 more
Organizations: Dalian University of Technology · Shanghai Jiao Tong University · The Chinese University of Hong Kong, Shenzhen · University of Science and Technology of China · Beihang University · Sun Yat-sen University · University of Oxford
Multi-arm collaboration is becoming a core capability in embodied manipulation. Recent vision-language-action (VLA) models integrate perception, language, and control, but most represent language as a single global instruction and do not provide an explicit mechanism for assigning and composing arm-specific behaviors. This design limits transfer to collaboration patterns that differ from those observed during training. We present MA-VLA, a unified framework for multi-arm collaboration via atomic action assignment. MA-VLA decomposes cooperative behavior into mid-level atomic prompts and allocates them to individual arms, enabling explicit subgoal specification and compositional reuse across tasks. To reduce reliance on fixed execution roles, we introduce Arm Shuffle, a training-time permutation of the observation, state, and assigned atomic prompts for each arm. This permutation enforces role-agnostic instruction following and supports recomposition into unseen coordination patterns, which we term multi-arm compositional generalization. We also construct a benchmark in which test-time collaboration patterns are absent in training set. Across simulation and real-world evaluations, prior state-of-the-art VLAs largely fail under these unseen collaborations, while MA-VLA consistently succeeds. These results indicate that structured, per-arm atomic action assignment offers a practical route to scalable generalization in multi-arm embodied systems. Code, models, and data are available at https://github.com/zhangzaibin/future-robots
Figures & tables
Figure 1: Comparison with prior VLA systems. Previous VLAs typically execute a task from a single high-level instruction. MA-VLA instead decomposes the instruction into a sequence of interpretable mid-level atomic actions and assigns them to individual arms, which enables recomposition to handle unseen collaboration patterns.
Figure 2: Multi-arm compositional generalization: unseen collaboration patterns induced by new combinations of arm roles, spatial states, and atomic actions, beyond the patterns observed in training.
Figure 3: MA-VLA Pipeline. A high-level instruction is first decomposed into a sequence of atomic prompts. Each arm is represented by a tuple containing its atomic prompt, state, wrist-view observation, and ground-truth action (e.g., Arm1: , Arm2: , Arm3: ). During training, Arm Shuffle randomly permutes these arm tuples, preventing the model from binding behaviors to fixed arm identities and promoting role-agnostic coordination across diverse collaboration patterns.
Table 3: Results on unseen out-of-domain collaboration tasks. Arrows denote the required stacking order of cube colors. Notation: GBR=G → B → R; RGB=R → G → B; BRG=B → R → G.
Stack Bowls
Place Cubes
Pass Toys
Stack Cubes
Method
ID
OOD
ID
OOD
ID
OOD
ID
OOD
Pi0 [ 3 ]
9/20
0/20
12/20
0/20
3/20
0/20
5/20
0/20
MA-VLA
12/20
10/20
15/20
8/20
6/20
2/20
8/20
2/20
Table 4: Real-world results on dual-arm SO101. Each task is evaluated in-domain (ID) and compositional out-of-domain (OOD). OOD reconfigures the program: Stack Bowls/Stack Cubes use reversed stacking, Place Cubes uses reversed placement, and Pass Toys uses an alternative passing order.
Figure 4: Setup of the real-world dual-arm SO101 tasks. Top: global view. Bottom: wrist views (left: left-wrist camera; right: right-wrist camera).
Atom
Shuffle
View Dropout
Out-of-domain
In-domain
0.0%
48.0%
✓
0.0%
58.0%
✓
✓
7.3%
52.0%
✓
✓
✓
15.3%
53.0%
Table 5: Ablation Experiment of Components.
Figure 5: Effect of shuffle rate on in- and out-of-domain settings.
Method
Out-of-domain
In-domain
Avg.
VLA Number
Pi0
0.0%
48.0%
24.0%
1
Separate Pi0
0.0%
61.0%
30.5%
3
MA-VLA
15.3%
53.0%
34.2%
1
Table 6: Comparison between MA-VLA and Separate Model.
Figure 6: Qualitative rollouts of MA-VLA in RoboFactory and on SO101 under in-domain and out-of-domain splits. Out-of-domain evaluates compositional generalization by requiring unseen recombinations of atomic actions under novel arm and object states.