Researchers have developed a method for improving the perception of underwater robots in degraded visual conditions. They used a pre-trained foundation model as the visual encoder and combined it with sonar information to adapt to changing conditions. The team created a benchmark with five levels of degradation, from clean to extreme, and compared different methods for fusion and adaptation. Their approach achieved a 33.5% relative improvement in accuracy over the baseline method under extreme conditions. The study shows that while pre-trained models can be valuable, they are insufficient on their own when dealing with severe information loss, and that adapting fusion mechanisms to modality reliability can improve robust underwater perception. --- Why it matters: This research matters because it addresses a significant challenge in underwater robotics: reliable perception in degraded visual conditions. By improving the ability of robots to perceive their environment under such conditions, this work has implications for applications like ocean exploration, inspection, and maintenance. Source: https://arxiv.org/abs/2608.19710