Multiview Progress Prediction of Robot Activities

A team of researchers from the University of Florence and University of Verona has published a paper titled 'Multiview Progress Prediction of Robot Activities,' accepted at ICASSP 2026. The work addresses a critical gap in robotics: enabling robots to understand the progress of their own actions in real-time. This capability, known as action progress prediction, is essential for robots to operate safely alongside humans, provide timely assistance, and make autonomous decisions. The researchers argue that this area has been largely overlooked, with single-camera systems often failing due to self-occlusion—where the robot's own body blocks its view of the task.

The proposed solution is a novel multi-view architecture that integrates data from multiple cameras to create a robust understanding of action progression. By fusing these perspectives, the system can accurately predict how far along a manipulation task is, even when parts are hidden from any single view. The architecture was validated through experiments on the Mobile ALOHA robot, a popular platform for bimanual mobile manipulation. This technical advancement moves beyond simple task completion detection to continuous progress estimation, a foundational step for more fluid human-robot collaboration and complex, multi-step autonomous operations.