Research map / Vision-language and multimodal / Robot action and navigation
AVA-VLA: Improving Vision-Language-Action models with Active Visual Attention
Lei Xiao, Jifeng Li, Juntao Gao, Feiyang Ye, Yan Jin, Jingjing Qian, Jing Zhang, Yong Wu, Xiaoyuan Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add AVA-VLA (recurrent state + active visual attention) | OpenVLA-OFT · LIBERO (one policy for all 4 suites) | success rate (%): 96.8 -> 98.0 | Improved |
| add AVA-VLA | OpenVLA-OFT · CALVIN ABC->D | average length: 4.28 -> 4.65 | Improved |
| add AVA-VLA | OpenVLA-OFT · LIBERO+ (one policy per suite) | success rate (%): 69.6 -> 74.7 | Improved |
| add AVA-VLA on non-robot-pretrained LLaMA2-7B backbone | OpenVLA-OFT · LIBERO-Long (single-task) | success rate (%): 90.0 -> 92.6 | Improved |
| add AVA-VLA on Qwen2.5-0.5B backbone | OpenVLA-OFT · LIBERO-Long (single-task) | success rate (%): 89.4 -> 90.8 | Improved |
| add state-based initialization to OpenVLA-OFT (ablation) | OpenVLA-OFT · LIBERO (one policy for all 4 suites) | success rate (%): 96.8 -> 97.5 | Improved |
| add AVA module to state-based initialization (ablation) | AVA-VLA · LIBERO (one policy for all 4 suites) | success rate (%): 97.5 -> 98.0 | Improved |
| add L_omega L2 regularizer (ablation) | AVA-VLA · LIBERO (one policy for all 4 suites) | success rate (%): 97.5 -> 98.0 | Improved |
| add AVA module to state-based initialization on CALVIN (ablation) | AVA-VLA · CALVIN ABC->D | average length: 4.63 -> 4.65 | Improved |
| prune 50% of visual tokens using AVA weights (ablation) | AVA-VLA · LIBERO (one policy for all 4 suites) | success rate (%): 98.0 -> 97.3 | Hurt |
| prune 90% of visual tokens using AVA weights (ablation) | AVA-VLA · LIBERO (one policy for all 4 suites) | success rate (%): 98.0 -> 93.9 | Hurt |
| train baseline and AVA-VLA under matched compute | AVA-VLA vs OpenVLA-OFT · LIBERO matched training (100K steps, batch 256) | success rate (%): 96.8 -> 98.3 | Improved |
Closest papers by meaning
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.93)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.92)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.92)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
