<![CDATA[The increasing deployment of artificial intelligence in robotic perception systems necessitates models that are both accurate and interpretable to ensure reliable decision-making in dynamic environments. This study proposes an intrinsically interpretable deep vision framework designed to enhance robustness and transparency in robotic perception tasks. The framework integrates convolutional feature extraction with embedded attention mechanisms, producing predictive outputs alongside spatially interpretable explanations. Experiments were conducted on publicly available benchmark datasets, including RGB-D Object Dataset, KITTI Vision Benchmark Suite, and adapted COCO subsets, covering scenarios with varying illumination, occlusion, and background complexity. Performance was evaluated through classification accuracy, precision, recall, localization consistency, and stability across repeated executions, with statistical validation using paired two-tailed t-tests and confidence interval analysis. Results indicate that the proposed framework maintains competitive accuracy while providing superior localization consistency, reduced variance, and stable attention behavior compared with conventional CNN baselines and post-hoc explanation methods. These findings demonstrate that embedding interpretability within the model architecture improves both predictive reliability and operational transparency. The proposed approach addresses key challenges in real-world robotic applications, facilitating safer automation, enhanced user trust, and alignment with regulatory expectations for explainable AI. By combining accuracy, robustness, and interpretability, this framework provides a scalable solution for intelligent robotic perception systems, supporting sustainable and responsible deployment in complex environments. The study highlights the critical role of intrinsic interpretability as a design principle for AI-driven robotics, offering practical insights for researchers, system developers, and policymakers seeking to advance trustworthy autonomous systems.]]>
