PTZJEPA: Predictive Active Vision for Autonomous Visual Data Collection

PTZJEPA is a closed-loop active perception framework that combines a predictive world model with reinforcement-learning control of a pan-tilt-zoom camera. The goal is to give visual sensing systems agency: instead of passively processing a fixed stream of images, the camera learns where to look next and adapts its data acquisition strategy to the visual environment.

The project integrates image-based joint-embedding predictive learning (I-JEPA) with reinforcement-learning control inspired by Dreamer. The agent learns predictive latent dynamics and uses them to choose camera movements that acquire informative observations. This links perception, prediction, and action in a single autonomous visual data collection loop.

Applications include autonomous high-resolution scanning, environmental sensing, edge AI, robotic perception, and other domains where an embodied sensor must collect useful data under bandwidth, compute, or field-of-view constraints.