In this study, we introduce an innovative approach to enhance interpretability in the design optimization of voxel-based soft robots (VSRs). VSRs present a unique challenge in achieving optimal designs due to their vast design space and the intricate relationships between individual voxels, compounded by the difficulty in interpreting the design choices and their functional implications. Traditional research has focused on meticulously adjusting grid voxels to optimize designs. However, this direct exploration of the vast design space often results in inconsistent outcomes and poses significant challenges to interpretation. To address these issues, we propose constraining the design space through part assembly and exploring designs using Bayesian optimization in a continuous feature space that offers higher interpretability. This approach facilitates diverse exploration and provides quantitative insights, marking a significant shift toward a more intuitive and exploratory design process in soft robotics. Our framework enables direct understanding of VSR designs, paving the way for future research and practical applications in this field.