The VersaCamVLA controller completed 58.3% of runs at camera positions absent from its training, researchers report. A standard version of its underlying model completed 13.3% under the same condition.
The physical test used a two-armed robot on three tasks: moving a cube into a bowl, stacking cubes and handing a test tube from one arm to the other for insertion. The researchers say each task had 20 evaluation runs per condition, for 60 runs per system at the new camera positions.
Vision-language-action (VLA) systems combine camera images with a written instruction to generate robot movements. Many are trained with a fixed set of camera views. Moving a camera can change the visual input the system relies on, even when the task stays the same.
VersaCamVLA combines its camera images into a compact set of numbers describing the scene. During training, it predicts how the scene would look from other camera angles, including object regions and edges, and uses different combinations of fixed and wrist-mounted cameras. As the robot moves, its wrist cameras add changing viewpoints without requiring extra demonstrations at each camera setup.
At familiar camera positions, the standard π0.5 model averaged 41.7% success and VersaCamVLA averaged 53.3%. At positions absent from training, the averages were 13.3% and 58.3%, respectively. In a separate test, the same VersaCamVLA model averaged 65% with five camera views and 53.3% with four; the standard model was not tested with five, so that result has no like-for-like baseline.
The evaluation remains small: one dual-arm platform, three tabletop tasks and 20 runs per task in each condition. The method also needs camera locations and calibration values; it does not select camera positions on its own. The paper shows how this approach handled the tested camera changes in a lab, not whether it will work reliably on other robots, tasks or operating environments.
Sources
- Han et al., “VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation,” arXiv, submitted Oct. 8, 2026
- Illustrative robot-arm photo by Shixart1985 (Wikimedia Commons metadata credits Nenad Stojković), licensed CC BY 2.0; Commons file page. It does not depict the study’s robot.
