MaskVLA
IROS 2026 Accepted

MaskVLA: Visual Masking Against Trajectory Overfitting of Vision-Language-Action Model

Yuxuan Jiang1,* Jiaying Huang1,2,* Ge Wang1,2,3,* Shenhao Yan1,3 Jiahao Yang1,3,4 Chengsi Yao1,3,5 Qi Liu1,3 Qing Zhao1,3 Shuguang Cui1,2 Yiming Zhao3 Yatong Han3,† Zhen Li1,2,†

* Equal contribution. † Corresponding authors.

1FNii, Shenzhen 2CUHK-Shenzhen 3Ising AI 4University of Glasgow 5Southeast University
Paper Code
79.8% average success on RoboTwin2.0 across easy and hard settings
+20.0 absolute hard-mode gain over OpenVLA-OFT
85.0% real-world success on Feed carrot and Place orange

Abstract

Reducing trajectory overfitting with view-balanced training.

Vision-Language-Action (VLA) models integrate vision-language understanding with executable robot actions, enabling end-to-end learning for robot control. However, when fine-tuned on limited manipulation datasets, existing models can overfit to demonstration trajectories and rely too heavily on the primary camera view.

MaskVLA introduces a simple masking-based fine-tuning strategy: randomly mask a small portion of the primary camera input during training so the policy learns to use wrist-camera information more effectively. On RoboTwin2.0, MaskVLA improves average success by 23.2% over π0 and 16.8% over OpenVLA-OFT, with additional gains on real-world ALOHA tasks.

Method

Random masking on the primary camera view.

The primary view gives global scene context, while wrist cameras provide close-range evidence during occlusion and contact. MaskVLA explicitly discourages primary-view shortcuts during fine-tuning.

MaskVLA architecture diagram
During training, a small portion of the primary camera input is masked. During inference, the policy receives full multi-camera observations.
01

Diagnose overfitting

Grad-CAM analysis shows baseline attention drifting toward local texture cues and primary-view bias.

02

Mask the shortcut

Primary-camera masking forces the policy to learn from all available visual streams.

03

Recover robustness

At inference time the mask is disabled, letting the model use full multi-view observations.

Results

Consistent gains in simulation and on ALOHA robots.

MaskVLA overview and results summary
Overview of the MaskVLA training strategy and evaluation settings.

RoboTwin2.0 benchmark

MaskVLA achieves the best average performance across four bimanual manipulation tasks. In hard mode, it reaches 75.0% average success, a 20.0-point gain over OpenVLA-OFT.

  • Best performance on 6 of 8 simulation subtasks.
  • Large gains on position-generalization tasks such as Stack bowls three.
  • MaskVLA-main-10% is the strongest ablation setting, averaging 74.0% in hard mode.
Real-world task success rates
Real-world ALOHA evaluation across Feed carrot, Place orange, Adjust cups, and Stack bowls.

Real-world ALOHA evaluation

On the Airbot Mobile ALOHA platform, MaskVLA improves over OpenVLA-OFT across all four real-world tasks, including 85.0% success on Feed carrot and Place orange and 70.0% on Adjust cups.

ALOHA real-world task setup

Videos

Qualitative rollouts.

Simulation comparisons use randomized hard settings. Real-world videos show successful ALOHA rollouts.

Simulation: MaskVLA success vs. baseline failure

Adjust Bottle

MaskVLA
Baseline

Open Laptop

MaskVLA
Baseline

Put Object Cabinet

MaskVLA
Baseline

Stack Bowls Three

MaskVLA
Baseline

Real-world ALOHA rollouts

Feed Carrot

Adjust Cups

Place Orange

Stack Bowls

Citation

@inproceedings{jiang2026maskvla,
  title  = {MaskVLA: Visual Masking Against Trajectory Overfitting of Vision-Language-Action Model},
  author = {Jiang, Yuxuan and Huang, Jiaying and Wang, Ge and Yan, Shenhao and Yang, Jiahao and Yao, Chengsi and Liu, Qi and Zhao, Qing and Cui, Shuguang and Zhao, Yiming and Han, Yatong and Li, Zhen},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year = {2026}
}