Situated Mapping for Transfer Learning

In transferring a task model learned in a source environment to a new, target environment, the agent may encounter novel objects. Thus, analogical mapping between objects in the source and target environments is both important and complex. Two objects that share the same purpose in one task may be mapped to different objects in the context of another task. We address this context- dependent object mapping by leveraging a human teacher. The teacher provides a limited number of object correspondences from the ground-truth object mapping, from which the remainder of the object mapping can be predicted. We evaluate this Mapping by Demonstration approach both in simulation and on two physical task examples (sorting and assembly). Our results show the agent can use human guidance to quickly infer a correct object mapping, requiring assistance with only the first 1/14 and 5/12 steps of the sorting and assembly tasks, respectively.