Multimodal systems typically assume all sensors are available, but in practice inputs can be lost. A new arXiv preprint argues that predictive video representations do not by themselves tell a system how to behave when a sensor is missing. The authors propose transferring frozen V-JEPA2 features to a heterogeneous sensor setup that includes RGB video, depth video, and other modalities.

The approach is evaluated for breathing activity and pose recognition. By keeping the V-JEPA2 features frozen, the method may offer a lightweight way to adapt a pretrained representation to multiple sensor types without retraining the entire model. This could make the system more robust to sensor dropout in real-world deployments.

The source is a single arXiv abstract, so details on the architecture, dataset, and quantitative results are not yet available. The preprint is listed as arXiv:2610.15891v1.