Multi-view vision-language-action (VLA) models still face challenges in spatial generalization, particularly under changes in camera viewpoints and robot initial states. These models typically receive action supervision only in the robot base frame, requiring them to implicitly learn the spatial correspondence between observations from each view and base-frame actions. We propose One View, One Action (OVOA-VLA), which learns to predict actions in each view's camera frame, allowing the same physical motion to provide diverse, observation-aligned supervision across views. OVOA-VLA jointly trains a base-frame expert and view-specific action experts, whose robot-state inputs and action targets are expressed in their respective camera frames. Two complementary directed attention graphs model both directions of interaction between the base and view-specific experts, jointly leveraging view-specific action supervision to improve base-frame prediction. Cross Graph Consistency further encourages agreement between the two graphs' base-frame predictions for the same sample. With the same multi-view inputs, the model supports both full and base-only inference. Experiments across multiple simulation benchmarks demonstrate improved manipulation performance and robustness to camera-viewpoint and robot-initialization perturbations. Full and base-only inference achieve average success rates of 88.1% and 85.7% on LIBERO-Plus, respectively. In real-world evaluations, base-only inference raises the average success rates from π0.5's 28.3% to 58.3% across three tasks under spatial distribution shifts.
Figures & tables
Figure 1: Motivation and comparison of action supervision. (a) The same physical action has different representations in the robot base frame and different camera frames; as the camera pose changes, its representation in the corresponding camera frame changes accordingly. (b) Existing multi-view VLAs typically supervise actions only in the base frame. OVOA-VLA retains this supervision and adds action supervision in each view’s camera frame using the same demonstration.
Figure 2: Overview of OVOA-VLA. (a) Each view-specific action expert attends to its own view and predicts camera-frame actions, jointly trained with the base expert. (b) Complementary attention graphs enable directed interactions: view-specific experts attend to base action features ( B→R ), while the base expert attends to view-specific action features ( R→B ). CGC encourages consistency between the two base-frame predictions. (c) Full Inference uses R→B ; Base-only Inference uses B→R , omitting view-specific experts while retaining all input views. Both execute base-frame actions.
Model
LIBERO
LIBERO-Plus
Spatial
Object
Goal
Long
Avg.
Camera
Robot
Total
UniVLA [ 6 ]
96.5
96.8
95.6
92.0
95.2
1.8
46.2
42.9
WorldVLA [ 8 ]
85.6
89.0
82.6
59.0
79.1
0.1
27.9
25.0
π0 -FAST [ 31 ]
96.4
96.8
88.6
60.2
85.5
65.1
21.6
61.6
OpenVLA-OFT [ 18 ]
97.6
98.4
97.9
94.5
97.1
56.4
31.9
69.6
DreamVLA [ 44 ]
97.5
94.0
89.5
89.5
92.6
65.0
40.9
69.9
Table 1: Performance comparison on LIBERO and LIBERO-Plus. We report success rates (%) across the four LIBERO task suites and their average. For LIBERO-Plus, we report Camera and Robot perturbation results and the overall success rate across all 10,030 evaluation instances. ∗ denotes results reproduced by us. Complete results for all perturbation dimensions are provided in Appendix C.3 .
Method
Avg. SR ↑
Avg. PS ↑
Avg. IS ↑
π0
47.0
–
–
π0 -FAST
51.2
–
–
π0.5∗
42.2
57.1
70.4
Ours (Base-only)
56.8
69.8
74.6
Ours (Full)
60.6
74.2
77.6
Table 2: Results on the VLABench Track-1 in-distribution benchmark. We report the average success rate (SR), progress score (PS), and intention score (IS) over 10 tasks. ∗ denotes results reproduced by us. Complete results are provided in Appendix C.5 .
Method
Avg. SR (%) ↑
π0.5∗
64.4
Ours (Base-only)
72.7 (+8.3)
Ours (Full)
75.3 (+10.9)
Table 3: Results on MimicGen Panda10 under the D0 setting. We report the average success rate (SR) over 10 tasks with 100 episodes per task. Complete results are provided in Appendix C.5 .
Method
LIBERO Avg.
Camera
Robot
Baseline ( π0.5 )
92.6
66.2
67.2
+ View-Specific Action Experts
96.5
73.1
71.6
+ Ref.-Aligned Visual Cond.
96.4
75.7
79.6
B→R Only
96.6
79.1
77.0
R→B Only
98.1
70.0
85.4
Bidirectional Attention
95.2
77.1
71.7
Table 4: Method design ablations on LIBERO and LIBERO-Plus. We report success rates (%); Camera and Robot denote the corresponding perturbations in LIBERO-Plus.
Mode
Avg. ↑
FLOPs ↓
Mem. ↓
Lat. ↓
Baseline ( π0.5 )
92.6
254.65
10.56
63.07
R→B
98.7
263.79
19.12
103.07
B→R
97.0
263.79
19.10
99.97
Base-only
96.3
254.66
10.49
62.77
Table 5: Performance and inference efficiency on LIBERO. We compare the average success rate (%), computational cost per action chunk (GFLOPs), peak GPU memory usage (GB), and mean inference latency (ms) across inference modes. Lower FLOPs, memory usage, and latency indicate better efficiency.
Figure 3: Real-world manipulation evaluation. Left: in-distribution success rates across six tasks, with 20 rollouts per task. Right: representative successful rollouts of the short-horizon Hang Mug task and the long-horizon Fill Drawer task.
Figure 4: Zero-shot robustness under spatial shifts. Left: task success rates pooled across the two shift types, with 20 rollouts per task. Middle and right: representative rollouts in which OVOA-VLA succeeds and π0.5 fails. Shift-specific results are shown in Figure 6 .
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Platform
Type
Dim.
Position
Orientation
Gripper
LIBERO / MimicGen
State
8
EEF position (3)
Rotation vector (3)
Finger joint positions (2)
Action
7
Translational increment command (3)
Rotation-vector increment command (3)
Gripper command (1)
VLABench
State
7
EEF position (3)
XYZ Euler angles (3)
Gripper state (1)
Action
7
Target position offset (3)
Component-wise XYZ Euler-angle difference (3)
Gripper command (1)
Real-world Franka
State
8
EEF position (3)
Quaternion in wxyz order (4)
Gripper closedness (1)
Action
7
Target position offset (3)
Relative rotation vector (3)
Target gripper closedness (1)
Appendix
Table 6: Proprioceptive state and action representations. Dimensions exclude padding. Numbers in parentheses indicate component dimensions. EEF denotes the end effector.
Method
Camera
Robot
Language
Light
Background
Noise
Layout
Total
UniVLA
1.8
46.2
69.6
69.0
81.0
21.2
31.9
42.9
WorldVLA
0.1
27.9
41.6
43.7
17.1
10.9
38.0
25.0
π0 -FAST
65.1
21.6
61.0
73.2
73.2
74.4
68.8
61.6
OpenVLA-OFT
56.4
31.9
79.5
88.7
93.3
75.8
74.2
69.6
DreamVLA
65.0
40.9
63.5
85.7
82.7
85.0
74.0
69.9
MemoryVLA
42.7
44.9
84.4
92.8
95.0
62.1
84.7
70.2
Appendix
Table 7: Performance comparison on LIBERO-Plus. We report success rates (%) under all seven perturbation dimensions, together with the overall success rate. ∗ denotes results reproduced by us.
Attention graph
Executed action
LIBERO
LIBERO-Plus
Spatial
Object
Goal
Long
Avg.
Camera
Robot
B→R
Camera
96.4
99.0
97.0
93.0
96.4
50.2
80.3
Wrist
86.6
99.4
90.8
56.6
83.4
70.0
58.8
R→B
Camera
98.4
99.4
96.2
96.8
97.7
47.8
84.6
Wrist
90.6
98.6
93.2
83.8
91.6
73.7
69.7
Appendix
Table 8: Success rates (%) when executing camera-frame and wrist-frame actions under the B→R and R→B attention graphs. Before execution, the predicted actions are transformed into the robot base frame using the corresponding camera extrinsics. Bold values indicate the better result between camera-frame and wrist-frame actions under the same attention graph for each perturbation type.
Method
Coffee
Coffee Prep.
Hammer
Kitchen
Mug
Square
Stack
Stack 3
Thread
3P Assembly
Avg.
π0.5∗
71
32
97
99
44
58
97
79
35
32
64.4
Ours (Base Only)
85
65
100
100
48
71
100
78
42
38
72.7
Ours (Full)
84
72
100
100
56
78
98
90
42
33
75.3
Appendix
Table 9: Performance comparison on MimicGen D0 distribution. We report the success rate (%) for each task and the average performance. ∗ denotes results reproduced by us.
Task
π0.5∗
Ours (Base Only)
Ours (Full)
SR
PS
IS
SR
PS
IS
SR
PS
IS
Add Condiment
54
75.3
98
50
72.7
98
66
80.7
96
Insert Flower
22
56
86
26
56
84
30
63
100
Select Book
50
56
76
64
73
88
74
82
98
Select Chemistry Tube
6
31
2
78
87
2
70
85
2
Select Drink
54
66
90
58
71
92
68
76
92
Appendix
Table 10: Task-wise performance on VLABench. We report Success Rate (SR), Progress Score (PS), and Intention Score (IS), with all values expressed as percentages. The Average row reports the unweighted mean across the ten tasks. ∗ denotes results reproduced by us.
Task
Language Instruction
Task Setup and Initialization
Success Criterion
Short-horizon tasks
Pick Cucumber
“Pick up the cucumber and place it into the blue bowl.”
A cucumber and two differently colored bowls are placed on the table. The robot must identify the blue bowl and place the cucumber inside it. During data collection, the cucumber’s initial position is randomized, and the positions of the two bowls are randomly exchanged.
After the gripper releases the cucumber, it lies entirely inside the blue bowl.
Stack Blocks
“Place the red block on top of the blue block.”
The robot must pick up the red block and stack it on the blue block. Both blocks’ initial positions are randomized without a fixed left–right ordering, with small variations in their in-plane orientations.
After release, most of the red block’s bottom face rests on the blue block, and the stack remains stable. Exact centering is not required.
Hang Mug
“Hang the mug on the mug rack.”
The robot must grasp the mug, adjust its orientation, and align its handle with a hook on the rack. The mug’s initial position and handle orientation are varied during data collection.
The handle is placed over the hook, and the mug remains stably supported by the rack after the gripper releases it.
Long-horizon tasks
Stack Three Cups
“Stack all the cups on the leftmost cup.”
Three equally sized cups of different colors are placed separately on the table. The robot must sequentially stack the other two cups onto the initially leftmost cup. Cup positions vary within their designated regions, with variations in relative placement and small orientation perturbations.
The initially leftmost cup remains at the bottom, and the other two cups are stacked above it, forming a single stable three-cup stack after release.
Appendix
Table 11: Real-world task definitions. We describe the language instruction, task setup, and success criterion for each short-horizon and long-horizon task.
Figure 5: Real-world setup. The training and additional test cameras are indicated.
Figure 6: Detailed zero-shot robustness under spatial distribution shifts. Success rates under camera viewpoint and robot initialization shifts. Each task aggregates two OOD configurations with five rollouts per configuration; Average denotes the mean across the three tasks.
Figure 7: Qualitative results under in-distribution evaluation. Representative successful OVOA-VLA rollouts across the six evaluated real-world tasks.
Figure 8: Qualitative results under robot initialization shifts. Representative successful OVOA-VLA rollouts on Stack Blocks, Erase Mark, and Stack Three Cups. The red box in the first frame of each rollout highlights the perturbed initial EEF pose.
Figure 9: Qualitative results on LIBERO. Successful execution trajectories from the Spatial, Object, Goal, and Long task suites.
Figure 10: Qualitative results on LIBERO-Plus. Each row shows a successful rollout under one perturbation category: Camera, Robot, Language, Light, Background, Noise, or Layout. The first column presents the initial state from the corresponding unperturbed rollout for comparison, while the remaining six columns show keyframes from the successful perturbed rollout, including its initial frame.
Figure 11: Qualitative results on VLABench. One successful execution trajectory for each of the ten evaluated tasks.
Figure 12: Qualitative results on MimicGen. One successful execution trajectory for each of the ten selected Panda D0 tasks.
We propose OC-VLA++, an extension of OC-VLA for viewpoint generalization under limited camera coverage. While OC-VLA grounds robot actions in the camera coordinate system to align action supervision with visual observations, camera-space grounding alone can still overfit to the few viewpoints observed during training. OC-VLA++ addresses this limitation by introducing geometry-guided paired-view supervision and an explicit cross-view action-equivariance objective. Given paired observations of the same manipulation scene from geometrically related viewpoints, the model is trained such that their camera-space predictions correspond to the same robot-frame action. This objective explicitly supervises how action predictions should transform across viewpoints, rather than relying solely on image-level augmentation. Experiments demonstrate substantial improvements in unseen-view generalization under limited camera coverage, with performance degrading more gracefully under increasing camera displacement. These results establish cross-view action equivariance as an effective complement to observation-centric action grounding for robust real-world deployment.
Tianyi Zhang, Ziyang Gong, Zhenjie Yang +2
Zhejiang University · Shanghai Jiao Tong University · The University of Hong Kong +2
Vision-Language-Action (VLA) models have shown remarkable promise in generalized robotic manipulation. However, their spatial generalization remains fragile. We argue that simply increasing the number of viewpoints is insufficient. Models often fall into the trap of Shortcut Learning, latching onto spurious correlations (e.g., fixed relative poses between objects or between the camera and robot base) rather than learning true spatial relationships. In this work, we propose a data-centric solution to enhance VLA spatial generalization. We utilize a dual-arm setup where one arm performs manipulation while the other serves as a mobile environmental camera. We systematically evaluate three data distribution patterns: Fixed, Multi-Fixed, and Moving Views. Our findings reveal that a hybrid strategy, combining continuous camera motion with diverse static viewpoints, yields the best performance by substantially reducing spurious correlations while maintaining training stability. Our experiments demonstrate that this strategy mitigates spurious correlations, enabling VLAs to generalize to unseen camera poses and object configurations where simply adding more static viewpoints fails. Crucially, we reveal that the susceptibility to shortcut learning and the struggle with spatial generalization are universal characteristics shared across diverse architectures. Consequently, all evaluated models (ACT, Diffusion, and VLA models including Pi0 and Gr00t) benefit significantly from our mixed data strategy.
Jincheng Tang, Yilong Zhu, Zhengyuan Xie +2
Lion Rock AI Lab, China Merchants Research Institute of Advanced Technology. · The Hong Kong University of Science and Technology. · Nankai University.
Occluded tasks remain a bottleneck in robot manipulation. Existing solutions either deploy additional physical cameras requiring training-inference camera parity, or rely on explicit 3D reconstruction with high computational cost. Moreover, both approaches rely on standard agent-view and wrist-view observations, while failing to capture occlusion information and future scene evolution. To this end, we propose UniviewVLA, a unified multiview Vision-Language-Action model with world modeling, which infers multiview scene evolution for action prediction from only standard two-camera observations. We demonstrate that by leveraging generated multiview future views from the world model, UniviewVLA reveals occluded cues and models future scene evolution, improving action prediction and removing the need for extra hardware or explicit reconstruction. Besides, to accelerate inference while preserving prediction accuracy, UniviewVLA develops Motion-Informative Token Compression, which compresses each generated view from 625 to 16 tokens and reduces per-view latency from 6-7s to 0.2-0.3s. UniviewVLA also proposes training-free Action-Entropy View Selection, which dynamically identifies the most action-informative view at different inference stages. Extensive experiments show that UniviewVLA achieves 95.8% on LIBERO and 4.60 on CALVIN ABCD to D, both standard occlusion-free benchmarks. On customized occlusion-focused tasks, it improves success rate from 40.0% to 73.3%, and average real-robot success rate by 33.4 points, demonstrating stronger occlusion-focused performance without sacrificing standard occlusion-free benchmarks.
Tao Xu, Runhao Zhang, Zhijian Huang +7
1Tongji University · 2Shanghai Innovation Institute · 4Xiaomi EV +1