LOOPER: A framework for synthetic dataset generation with configurable sensors and multi-view XR environments

Collecting multi-view datasets is essential for training and evaluating AI models in domains such as robotics and autonomous systems. However, generating such datasets remains challenging due to sensor synchronization issues and the labeling process, which is often time-consuming, error-prone, and dependent on manual intervention. To address these limitations, a novel framework, LOOPER (Light Object and event Oriented Package recordER), is proposed, enabling the automatic generation of labeled dataset based on existing XR applications. This is achieved through a two-step procedure: (i) trajectory and event recording, and (ii) multiple sensors data collection. This approach enables the creation of reproducible, multi-view datasets while simplifying the data generation pipeline. Experiments demonstrate that LOOPER (i) doesn’t compromise the user experience, (i) enables the synchronized generation of various data types (RGB, Depth, LiDAR, Human Pose) and (iii) is usable in several contexts including human action recognition and robotic perception.