Figure 7
Training and testing of the observer and corresponding success rate We first gathered training and testing data by randomly placing the actor, two green food items, and the obstacle (Table, A). We also collected “obscured” test cases where we deliberately placed the closest food to where it is not visible to the actor (Table, B). Higher success rates were achieved by balancing the training data with half “visible” data and half “obscured” data (Table, C and D). Learning curves across all four above scenarios are shown. Error bars are presented to show experiment results across three different random seeds used in both data splitting and network training.
