15825aee15eb335cc13f9b559f166ee8-AuthorFeedback.pdf

Taking a knife/apple is the same...# affordances is very low. The reward in Eqn 1 is provided for every new10 interaction executed by the agent, where an interaction consists of an action (take, slice) coupled with an object11 instance(apple1,lettuce)--i.e. Impact on environment if one tries every action/object. Weuse the true object boundaries from the simulator. Interleaving the affordance training more32 frequently results in better segmentation (+1.8% mAP) but no improvement in interaction exploration.