Hierarchical Task-Parameterized Learning from Demonstration for Collaborative Object Movement

Figure 8

Comparison of generalization errors when different numbers of situations were demonstrated. The robot behavior abbreviations shown at the top stand for vanilla TP-GMM (VT) and hierarchical structure with TP- GMM (HS). Top panels show task configurations and samples of task situations, where each task frame was sampled from the specified range. When ten situations were demonstrated, the via frame of the tenth situation was drawn from an outlier range. Panels in the bottom three rows show boxplots of error metrics with different numbers of demonstrated situations in the corresponding task configurations. The center box lines represent the medians, and the box edges are the 25th and 75th percentiles. Circles show mean values. An asterisk and a horizontal line below boxplots show that the mean error from HS is significantly lower than the mean error from VT with .

(a)

(b)

(c)