Researchers at Shanghai Jiao Tong University developed C³-JEPA, a model that uses an underwater salvage robot’s cameras and control signals to predict how its target and gripper will move. In two real underwater recordings, it reconstructed object features for a camera view omitted from its input, but the tests replayed recorded commands rather than letting the model steer the vehicle.
During a close salvage approach, a work-class remotely operated vehicle (ROV) may see only part of a target from any one camera, while thruster commands take time to change the vehicle’s motion. C³-JEPA receives synchronized color-camera views and thruster and gripper commands, with no contact sensors.
Instead of generating future video, the model keeps separate representations for the target and gripper, combines evidence across camera views, and predicts how those representations will change. During training, it hides one camera’s view and tests whether the other views can supply the missing object information. This encourages its internal state to retain a target’s features when the target leaves one camera’s view.
The predictive interface can rank candidate thruster commands by their forecasted outcomes. The authors connected it to a conventional controller in simulation, where the combined system completed an approach, alignment, contact and hold sequence; they did not report a closed-loop field trial.
The real-video evaluation used two deployments in a test pool, one tracking a submerged steel frame and one bringing the gripper into contact with an aluminum pipe. In a 36-step forecast that fed each prediction into the next and used recorded commands, the model’s error in its encoded scene state was 30% lower than a baseline that repeated the previous state. This compares learned visual features with features from recorded video, not distance to a real target: the cameras were uncalibrated and the team had no metric ground truth. A separate image-mask check—how closely predicted target and gripper regions overlapped the regions in the reference video—showed no clear gain over persistence.
The result supports prediction from these two underwater recordings, not autonomous salvage. The arXiv preprint says the work has been submitted to IEEE for possible publication and that code and data are available from the authors on request.
Sources
- Yang et al., “Underwater C³-JEPA: An Object-Centric Cross-View World Model for ROV Salvage,” arXiv:2609.30214v1 (submitted September 24, 2026)
- Lead image: NOAA’s Jason ROV collects a sample near West Mata Volcano; NOAA/National Science Foundation, U.S. public domain. This is an illustrative ROV photograph, not the vehicle used in the study. Image and rights record.