Newest papers

FocusVLA: Future-Intent Guided Dynamic View Routing for Vision-Language-Action Models

Huixin Xu, Ziya Yu, Peng Li · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.31423 · cited by 0 works (OpenAlex) · open-access PDF

Indonesian

Karena manipulasi yang dapat diandalkan membutuhkan pengamatan yang memadai dari objek, efek akhir, dan hubungan spasial mereka, input multi-view secara luas digunakan untuk memberikan informasi spasial komplementar. Namun, eksperimen kami menunjukkan bahwa model VLA yang ada tidak sepenuhnya memanfaatkan komplementaritas multi-view inti. Kebijakan ini cenderung untuk terus bergantung pada tampilan pergelangan tangan, tidak memiliki kemampuan untuk secara dinamis menyesuaikan ketergantungan tampilan global-lokal di seluruh tahap manipulasi, yang membatasi adaptabilitas stadium dan ketahanan. Untuk mengatasi masalah ini, kami mengusulkan FocusVLA, framework routing multi-view yang sadar stadium. FocusVLA memprediksi niat interaksi masa depan dari tindakan historis dan menggunakannya untuk membimbing tampilan dinamis selama generasi tindakan. Dengan beradaptasi meningkatkan koordinasi global dan pengamatan global di

Machine translation (Meta M2M100); the English original follows.

Abstract

Vision-language-action (VLA) models have become an important paradigm for robotic manipulation by unifying visual perception, language understanding, and action generation. Since reliable manipulation requires sufficient observation of objects, end-effectors, and their spatial relationships, multi-view inputs are widely used to provide complementary spatial information. However, our experiments show that existing VLA models do not fully exploit such multi-view complementarity. The policy tends to persistently rely on the wrist view, lacking the ability to dynamically adjust global-local view dependency across manipulation stages, which limits stage adaptability and robustness. To address this issue, we propose FocusVLA, a stage-aware multi-view routing framework. FocusVLA predicts future interaction intent from historical actions and uses it to guide dynamic view routing during action generation. By adaptively coordinating global and local observations across different task stages, FocusVLA improves multi-view utilization. On RoboTwin 2.0, FocusVLA improves the average success rate over the strongest baseline $π_{0.5}$ by 10.10 pp. In real-world evaluation, FocusVLA improves the success rate from 71.7\% to 85.0\%, achieving a 13.3\% gain over $π_{0.5}$. Moreover, the intent prediction and dynamic routing modules of FocusVLA can be integrated into existing VLA frameworks without modifying their core architectures.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Exploit, Adaptability, Task (project management), Routing (electronic design automation), Dependency (UML)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange