Figure 5
The diagram is a horizontal sequence of four tall, colored columns with rounded labels at the top, connected left to right by thick black arrows to show process flow. The first dark blue column is headed “H I M P” and describes a “proposed hybrid deep learning intention prediction model,” listing three contributing components in stacked text: “C N N: visual features”, “L S T M: temporal dependencies”, and “CLIP: semantic understanding”. The second block, in bright turquoise, has a top circle labeled “S H A P”. Its heading reads “SHAP (SHapley Additive Explanations)” followed by bullet points listing: “Feature importance analysis”, “Positive and negative contributions”, “Consistent explanations”, and “Quantifies the influence of U T M input features and Explains model predictions to human workers”. The third block is a gray vertical rectangle with a circle at the top labeled “Grad-C A M”. Its heading reads “Grad-C A M (Gradient-weighted Class Activation Mapping)”, and below are bullet points: “Provides heatmaps that show areas in an image that the model finds most relevant” and “Improves human-robot collaboration by helping maintenance teams understand why the model flagged certain areas as needing attention”. On the far right, a green vertical rectangle with a circle labeled “Visual tools” at the top serves as the final step. Its heading reads “Real-time user-friendly visualization tools”, followed by bullet points: “Integrate with S H A P and Grad-C A M”, “Real-time highlights the most important regions that contributed to an AI model’s decision”, and “Dashboards for communicate uncertainties visually to improve trust”.Roadmap for future work regarding interpretability