CVPR 2026

AVA-VLA: Improving Vision-Language-Action models with Active Visual Attention

Lei Xiao1,*, Jifeng Li1,*, Juntao Gao1,2, Feiyang Ye1,†, Yan Jin1, Jingjing Qian3, Jing Zhang2, Yong Wu1, Xiaoyuan Yu1,†

1Li Auto Inc.    2Beijing University of Technology    3The Chinese University of Hong Kong, Shenzhen

*Equal contribution    †Corresponding authors

Paper Code arXiv
Figure 1 for AVA-VLA

Figure 1. (a) Visualized comparison of the proposed AVA-VLA framework and vanilla VLAs. (b) Qualitative comparison of visual focus from two viewpoints while executing the task “turn on the stove and put the moka pot on it.” The vanilla OpenVLA-OFT [20] baseline fails to locate the task-critical “stove” switch, whereas AVA-VLA exhibits more stable focus by leveraging historical context.

Abstract

Vision-Language-Action (VLA) models have shown remarkable progress in recent embodied tasks, but most existing methods process visual observations independently at each timestep. This history-agnostic design effectively treats robot manipulation as a Markov Decision Process, even though real-world robotic control is inherently partially observable and requires reasoning over past interactions. To address this mismatch, we reformulate VLA policy learning from a Partially Observable Markov Decision Process perspective and propose AVA-VLA, a framework that conditions action generation on a recurrent state that serves as a neural approximation to the agent's belief over task history. Built on this recurrent state, we introduce Active Visual Attention (AVA), which dynamically reweights visual tokens in the current observation to focus on regions that are most relevant given both the instruction and execution history. Extensive experiments show that AVA-VLA achieves state-of-the-art performance on standard robotic benchmarks, including LIBERO and CALVIN, and transfers effectively to real-world dual-arm manipulation tasks. These results demonstrate the effectiveness of temporally grounded active visual processing for improving VLA performance in robotic sequential decision-making.

Highlights

POMDP Perspective for VLA

AVA-VLA reformulates vision-language-action policy learning from a partially observable perspective and uses a recurrent state to summarize task history for action generation.

Active Visual Attention

The AVA module dynamically reweights current visual tokens using both the textual instruction and execution history, improving attention to task-relevant regions.

Strong Empirical Results

The framework achieves strong results on LIBERO, CALVIN, and Mobile ALOHA real-world experiments, while also providing interpretable visual dynamics.

Method

Figure 2 placeholder for AVA-VLA framework overview

Figure 2. Overview of the proposed AVA-VLA framework. At each timestep, the recurrent state is projected from the previous hidden state to preserve historical context and to initialize the current action tokens. Then the AVA module combines this recurrent state with text-conditioned visual features from the current observation to generate soft importance scores, which modulate the visual attention matrices throughout the backbone LLM, enabling the model to focus on task-relevant regions based on both temporal context and current perception.

Main Results

Table 1 for LIBERO comparison

Table 1. Comparison on the LIBERO benchmark. The results are reported in two groups: one policy for all 4 suites, and one policy per suite. The best results in each column of each group are highlighted in bold.

Table 2 for CALVIN comparison

Table 2. Comparison on the CALVIN ABCD benchmark. The results are reported in terms of success rates (%) and average length. The best results in each column are highlighted in bold.

Figure 3 placeholder for Mobile ALOHA results

Figure 3. Comparison on the Mobile ALOHA real-world experiments. Evaluation across four manipulation tasks, including (a) Pick and Place, (b) Sequenced Instruction Understanding, (c) Flexible Object Folding, (d) Dexterous Action. Left: Representative middle states for each task setup. Right: Task-specific success rates and cross-task averages for our method and baselines.

Analysis

Table 4 for ablation study

Table 4. Ablation study on the two key components in the AVA-VLA framework. The results on the LIBERO benchmark in terms of success rates (%) under the “one policy for all 4 suites” setting are reported. The best results in each column are highlighted in bold.

Figure 4 placeholder for visual dynamics

Figure 4. Visual dynamics. The evolution of soft weights during the task “put both moka pots on the stove” from two viewpoints.

Table 5 for pruning ratio study

Table 5. Study on the visual token pruning with different pruning ratios. The results on the LIBERO in terms of success rates (%) under the “one policy for all 4 suites” setting are reported.

Qualitative Videos

Mobile ALOHA Real-World Tasks

Representative real-world dual-arm manipulation demos in the Mobile ALOHA experiments.

Scoop corn into bowl.

LIBERO Benchmark Tasks

Qualitative rollouts on representative LIBERO benchmark manipulation tasks.

Turn on the stove and put the moka pot on it.

BibTeX

@article{xiao2025ava,
  title={AVA-VLA: Improving Vision-Language-Action models with Active Visual Attention},
  author={Xiao, Lei and Li, Jifeng and Gao, Juntao and Ye, Feiyang and Jin, Yan and Qian, Jingjing and Zhang, Jing and Wu, Yong and Yu, Xiaoyuan},
  journal={arXiv preprint arXiv:2511.18960},
  year={2025}
}