Researchers use CLIP to enable 24-hour agricultural robots at night

Researchers have developed an unsupervised image translation framework that enables agricultural robots to navigate fields at night by converting daytime RGB images into realistic near-infrared (NIR) night scenes. The approach leverages a pre-trained Contrastive Language-Image Pre-training (CLIP) model to preserve semantic consistency during the day-to-night translation, eliminating the need for costly pixel-level annotations of nighttime data. A visibility mask accounts for the limited effective range of NIR illumination.

To validate the method, the team introduced AgriNight, a benchmark dataset of 428 daytime and 549 nighttime images collected from mobile robots in agricultural fields, all with pixel-wise semantic labels. The framework outperforms state-of-the-art image translation baselines in downstream semantic segmentation tasks. Real-time autonomous navigation experiments with a physical robot operating at night confirmed viability. Accepted to IROS2026, the work promises 24-hour crop monitoring, fruit harvesting, and nocturnal pest detection without manual annotation overhead.