Figure 8
Counterfactual perturbation and prediction sensitivity. (A) The first column shows the original observation of the observer and the third column shows the prediction from the observer. The second column is the real trajectory of the actor robot after releasing it. We then move the obstacle to a different location from the first column. The resulted new observation of the observer is shown in the fourth column. The last column shows the counterfactual prediction after seeing this new observation. Both examples show the observer changes its prediction after the physically closest green dot becomes visible from being obscured by the obstacle. (B) Our model is able to give accurate prediction when the sizes of different objects as well as actor robot is changed. However, the observer model performs worse when the color is changed by 40%, and eventually breaks when the color is changed by 60%. The size change in these experiments is limited by the arena size (128 \(\times \) 128 pixels).
