Underwater manipulator grasping remains challenging because image blur, light attenuation, and flow-induced disturbances degrade perception and control. These factors make target localization, contact judgment, and stable lifting difficult, especially when visual degradation and tactile fluctuation occur together. We propose AVT-TD3, an adaptive visual–tactile fusion reinforcement learning method for underwater manipulator grasping. AVT-TD3 constructs a unified policy state from visual observations, short-horizon tactile variations, and manipulator proprioception. A gated fusion module adjusts the contribution of each sensory branch, while an action modulation mechanism limits abrupt velocity-command changes during contact establishment and lifting. We train the continuous grasping policy with Twin Delayed Deep Deterministic Policy Gradient (TD3) and evaluate it in simulation under different turbidity, flow velocity, and target conditions, followed by controlled water-tank feasibility validation. Simulation results show that AVT-TD3 achieves better performance than Deep Deterministic Policy Gradient (DDPG), Soft Actor-Critic (SAC), and standard TD3 in success rate, completion steps, slip rate, and velocity-command smoothness. In the standard test scenario, AVT-TD3 achieves a success rate of 92.7%, an average of 76 completion steps, a slip rate of 4.1%, and an action variation magnitude of 0.20. Controlled water-tank tests further support the feasibility of deploying AVT-TD3, although open-water validation remains for future work.

Despite the ocean’s vast energy reserves and extensive coverage of Earth’s surface, the complexity of the underwater environment has hindered effective target recognition. To mitigate the feature degradation caused by underwater scattering, wavelength-dependent absorption, and non-uniform illumination, this study proposes a degradation-aware YOLOv11s-based detection framework for underwater robotic object detection. The framework enhances detection robustness by improving feature reconstruction, channel–spatial attention, and bounding-box regression in a unified architecture. First, to address limitations in model parameters, spatial channel reconstruction convolutions (SCConv) replace certain traditional convolutions. Sequential reconstruction via SRU-CRU effectively suppresses spatial and channel redundancy, enabling a more precise capture of underwater target deformations and complex features. Second, the Shuffle Attention module enhances the interaction between channel and spatial features, improving the model’s fine-grained representation of underwater targets, highlighting granular objects and key textures. Finally, Focaler-IoU is employed to linearly remap the IoU interval, improving the accuracy and convergence stability of bounding-box regression. These components work together to improve the model’s robustness in degraded underwater scenes. In the underwater robotic detection task, the improved model achieves an mAP@0.5 of 88.4%, which is 3.2 percentage points higher than that of the baseline YOLOv11s. These results indicate that the proposed model improves detection accuracy while maintaining the real-time requirements of underwater robotic applications.