Fig. 5

Phase two: subjective, user-focused metrics reveal differences that objective performance-based measures may not. a Completion times did not significantly differ for any of the tasks. The mKF was generally faster for tasks perceived as more difficult, and slower for tasks perceived as less difficult. b The mKF and CNN were significantly easier to use than the MLP for the fragile egg task, as evidenced by a lower NASA Task Load Index (TLX) score. c mKF and CNN outperform MLP in fragile egg task, although the difference was insignificant. All algorithms had high success in the pouring task. d All algorithms performed well and did not significantly differ in the zipping and folding tasks. e The CNN was significantly preferred over the MLP, aggregated across tasks. The mKF trended towards significance over the MLP across all tasks. The CNN was significantly preferred over the MLP for the fragile egg and zipper tasks. Boxplots show median (red line), inter-quartile range (colored boxes), and most extreme, non-outlier values (1.5 × inter-quartile range; black whiskers)