Research map / Vision-language and multimodal / Robot action and navigation
$\text{A}^3$-VLA: Automatic Perception-Guided Attention Alignment for Robot Manipulation
Shengzhe Zhang, Qi Zhang, Dazhong Shen, Chao Wang, Shaopeng Zhai, Hui Xiong
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes A^3-VLA, which aligns a VLA's attention with task-relevant regions (target objects and interaction areas) using automatically generated perception guidance, to counter reliance on task-irrelevant visual shortcuts.
Problem
Vision-language-action models trained only with action prediction lack fine-grained task-specific perceptual supervision, so they rely on task-irrelevant visual shortcuts and attend away from target objects and key interaction areas.
Approach
Automatic perception-guided attention alignment: supervises the VLA's attention toward target objects and interaction regions using automatically derived perceptual signals, rather than relying on action loss alone. Based on the abstract only; details unknown.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add language-to-image attention allocation loss (ablation) | A3-VLA (pi0) · VLABench (5 suites avg) ablation | success rate: 0.52 -> 0.64 | Improved |
| add state-to-image attention allocation loss (ablation) | A3-VLA (pi0) · VLABench (5 suites avg) ablation | success rate: 0.60 -> 0.64 | Improved |
| add contrastive language-target loss (ablation) | A3-VLA (pi0) · VLABench (5 suites avg) ablation | success rate: 0.61 -> 0.64 | Improved |
| add A3-VLA guidance to pi0 | pi0 · VLABench (5 suites avg) | success rate: 0.47 -> 0.64 | Improved |
| add A3-VLA guidance to OpenVLA-OFT | OpenVLA-OFT · VLABench (5 suites avg) | success rate: 0.16 -> 0.40 | Improved |
| add A3-VLA guidance to pi0.5 | pi0.5 · VLABench (5 suites avg) | success rate: 0.41 -> 0.67 | Improved |
| add A3-VLA guidance | pi0 / OpenVLA-OFT / pi0.5 · RoboTwin 2.0 (4 tasks avg) | success rate: 0.42 -> 0.52 (pi0) | Improved |
| add A3-VLA guidance to pi0 | pi0 · Real-world select bottles | success rate: 0.25 -> 0.75 | Improved |
| add A3-VLA guidance under texture perturbation | pi0.5 · VLABench TP | success rate: 0.18 -> 0.42 | Improved |
| add A3-VLA guidance under Gaussian noise | pi0.5 · VLABench NP | success rate: 0.33 -> 0.57 | Improved |
| increase attention-loss weight lambda1 or contrastive weight lambda2 (ablation) | A3-VLA (pi0) · VLABench | success rate | Hurt |
| raise attention threshold eta1 beyond moderate value (ablation) | A3-VLA (pi0) · VLABench | success rate | Hurt |
Closest papers by meaning
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- Object-Centric Prompt-Driven Vision-Language-Action Model for Robotic Manipulation (CVPR 2025 · similarity 0.91)
- FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.91)
- BridgeVLA: Input-Output Alignment for Efficient 3D Manipulation Learning with Vision-Language Models (NeurIPS 2025 · similarity 0.91)
- CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models (CVPR 2025 · similarity 0.91)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.90)
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.
