FocusVLA Teaches a Robot to Pick a Camera Based on Task Stage: +13.3 pp in Real-World Tests

A robot arm with several cameras has to see both the whole scene and what is happening right at the gripper. But according to researchers at the Technology Center of Software Engineering of ISCAS, current VLA models use these viewpoints in a very lopsided way. Their new system, FocusVLA, adds an average of 10.10 percentage points of success on RoboTwin 2.0 over the strongest baseline, π0.5, and lifts the result in real-world experiments from 71.7% to 85.0%. The work has been accepted at the CoRL 2026 conference.

A policy that only looks at the wrist

Vision-language-action models combine visual perception, language understanding, and action generation in a single policy. A reliable grasp requires observing the object, the end effector, and how they are positioned relative to each other. That is why robots carry several cameras at once: a global overview camera and a local one mounted on the wrist. The assumption is that the views complement each other.

The FocusVLA authors tested whether this holds in practice and got an unwelcome answer. Existing VLAs do not fully exploit the complementarity of the views. The policy stubbornly clings to the wrist camera and does not rebalance between the global and local views as the task progresses. Yet different stages of manipulation call for different information. While approaching an object, the overall picture of the scene is useful. Right before contact, details at the gripper matter more. Being locked onto a single viewpoint hurts stage adaptability and robustness.

Intent instead of a rigid scheme

FocusVLA is built as a stage-aware view-routing framework. The system takes the history of actions already performed and uses it to predict the future interaction intent, that is, what the robot is about to do next. This prediction drives dynamic view routing during action generation. The model decides for itself which viewpoint to emphasize at a given moment, and it reconciles global and local observations across the different stages of the task.

An important engineering detail: the intent-prediction module and the dynamic routing module can be embedded into existing VLA frameworks without changing their base architecture. This is closer to a plugin than to a new model built from scratch.

What the measurements show

Numbers from the paper:

• RoboTwin 2.0: average success is 10.10 pp higher than the strongest baseline, π0.5 • Real-world experiments: success rose from 71.7% to 85.0% • Difference from π0.5 in real-world tests: 13.3 pp • Acceptance: CoRL 2026 • Authors: Huixin Xu, Ziya Yu, Peng Li • Organization: Technology Center of Software Engineering of ISCAS • Preprint: arXiv 2610.31423

The gain on the real robot is even larger than in simulation. In a field where the gap between benchmark and physical world has traditionally been painful, that is an important signal. Still, keep in mind that the baseline is π0.5, one of the most prominent open VLA models, and FocusVLA wins specifically against it. Comparisons with other policies and a per-task breakdown remain questions for the full text of the paper.

Where this is useful in practice

The authors' observation applies to any team training manipulation policies on multiple cameras. If a model systematically ignores the overview view, adding a second or third camera raises the cost of the rig and the volume of data while bringing almost no benefit. A simple check of which view the policy relies on at different stages shows whether your system has the same problem.

The second practical takeaway concerns compatibility. Because the modules plug into existing VLAs without changing the core, you can try them on an already trained architecture without rebuilding the whole pipeline. For the warehouse and assembly robot industry, where stability across long, multi-stage operations matters more than records on individual demos, this logic looks workable.

For now, the paper's page lists no linked models, datasets, or demo spaces, so independent verification of the results depends on whether the authors release their code. But the idea itself, that a robot's intent determines where it should look, looks like a natural next step for VLAs with multi-camera inputs.