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

FocusVLA. Future-Intent Guided Dynamic View Routing for Vision-Language-Action Models. By Huixin Xu and two others, from the Nanjing Institute of Software Technology, the Institute of Software, Chinese Academy of Sciences, and three other institutions. Published on the twenty-second of October. No code repository is linked to it.

Vision-language-action models, VLA models for short, have become an important paradigm for robotic manipulation, joining perception, language understanding and action generation in one model. Reliable manipulation depends on seeing enough of the objects, the end-effector and how they sit relative to each other, so these models are commonly given several camera views that complement one another spatially.

The authors' experiments show that existing VLA models do not make full use of that complementarity. The policy keeps leaning on the wrist view. What it lacks is the ability to shift its dependence between the global view and the local one as manipulation moves from stage to stage, and that limits both how well it adapts to each stage and how robust it is.

FocusVLA is a stage-aware framework for routing between views. Its central move is to look ahead. From the actions taken so far, it predicts the interaction the robot is about to make, its future intent, and that predicted intent is what decides which views to rely on while the next actions are being generated.

In practice the routing is dynamic: as the task passes through different stages, FocusVLA adaptively coordinates global and local observations instead of fixing one view as dominant, and so makes better use of multiple views. The intent prediction module and the dynamic routing module are separable from the rest. They can be integrated into existing VLA frameworks without modifying the core architecture of those frameworks.

On RoboTwin two point zero, FocusVLA raises the average success rate over the strongest baseline, pi zero point five, by ten point one zero percentage points. In real-world evaluation, success rises from seventy-one point seven percent to eighty-five point zero percent, a gain of thirteen point three over pi zero point five.

Existing policies keep relying on the wrist camera, whatever stage a task has reached. FocusVLA lets an anticipated next interaction choose where the model looks, so its balance of global and local views changes as the task does.