Improving Grounded Natural Language Understanding through Human-Robot Dialog

摘要: Natural language understanding for robotics can require substantial domain- and platform-specific engineering. For example, mobile robots to pick-and-place objects in an environment satisfy human commands, we specify the humans use issue such connect concept words like red physical object properties. One way alleviate this engineering a new domain is enable environments adapt dynamically---continually learning constructions perceptual concepts. In work, present end-to-end pipeline translating natural commands discrete robot actions, clarification dialogs jointly improve parsing grounding. We train evaluate agent virtual setting on Amazon Mechanical Turk, transfer learned platform demonstrate it real world.