Underwater robotic vision encounters significant challenges, necessitating advanced solutions to enhance performance and adaptability. This paper presents MARS (Multi-Scale Adaptive Robotics Vision), a novel approach to underwater object detection tailored for diverse underwater scenarios. MARS integrates Residual Attention YOLOv3 with Domain-Adaptive Multi-Scale Attention (DAMSA) to enhance detection accuracy and adapt to different domains. During training, DAMSA introduces domain class-based attention, enabling the model to emphasize domain-specific features. Our comprehensive evaluation across various underwater datasets demonstrates MARS's performance. On the original dataset, MARS achieves a mean Average Precision (mAP) of 58.57\%, showcasing its proficiency in detecting critical underwater objects like echinus, starfish, holothurian, scallop, and waterweeds. This capability holds promise for applications in marine robotics, marine biology research, and environmental monitoring. Furthermore, MARS excels at mitigating domain shifts. On the augmented dataset, which incorporates all enhancements (+Domain +Residual+Channel Attention+Multi-Scale Attention), MARS achieves an mAP of 36.16\%. This result underscores its robustness and adaptability in recognizing objects and performing well across a range of underwater conditions. The source code for MARS is publicly available on GitHub at https://github.com/LyesSaadSaoud/MARS-Object-Detection/