Joydeep Biswas, University of Texas at Austin - Talk Abstrac

ABSTRACT:

We are interested in building and deploying service mobile robots to assist with arbitrary end-user tasks in everyday environments. In such open-world settings, how can we ensure that robots discern perceptual cues of relevance without requiring pre-enumeration of entities of interest and without prior knowledge of what users will ask of them? In this talk, I will survey these technical challenges, and present several promising directions to address them. To "get it right", robots will have to reason about open-world semantic- and geometric- perceptual cues, build unstructured long-horizon memories and revisit them as needed on-the-fly, and accurately infer specifications for novel tasks from limited human input.