Automated detection of feeding behavior in intensively farmed striped catfish (Pangasianodon hypophthalmus) is a prerequisite for data-driven feed management in Mekong-Delta aquaculture, where feed represents up to 75% of production cost and feeding is still assessed manually. The task is formulated as a world-model problem: rather than classifying isolated clips, the system learns a compact latent model of how the pond surface evolves over time. CTU-CatraWorld is proposed, a Vision-Language Joint-Embedding Predictive Architecture (VL-JEPA) whose latent dynamics are scored by an Energy-Based Model (EBM) head performing joint state classification and overfeeding-anomaly detection. Beyond the model, we contribute a rigorous, measured empirical study on a 45-video field corpus that isolates the core difficulty of the domain—the adversarial pair of active feeding (S1) versus overfeeding / residual feed (S2*), which are visually near-identical but demand opposite dispenser actions. We show that a single pellet-appearance cue cannot separate the pair (macro-F1 of 0.46), that raw optical-flow motion is confounded by hand-held camera egomotion (macro-F1 of 0.49), and that a frozen ImageNet ViT-B/16 backbone—the CTU-CatraWorld encoder prior to self-supervised adaptation—already reaches macro-F1 of 0.82 and AUROC of 0.92 on this pair, extending to macro-F1 of 0.67 over three states. The CTU-CatraWorld energy head trained on these features classifies the pair at macro-F1 of 0.76, and a latent-space JEPA adaptation of the features raises the pair to AUROC 0.98 (linear probe) and 0.96 (energy head), demonstrating the effectiveness of latent predictive adaptation. These measured results justify the architectural choices, quantify the separability of the challenging feeding states on field video, and establish a strong foundation for end-to-end self-supervised world-model learning in aquaculture.

View online