Real-Time Human-Robot Communication for Manipulation Tasks in Partially Observed Environments

In human teams, visual and auditory cues are often used to communicate information about the task and/or environment that may not otherwise be directly observable. Analogously, robots that primarily rely on visual sensors cannot directly observe some attributes of objects that may be necessary for reference resolution or task execution. The experiments in this paper address natural language interaction in human-robot teams for tasks where multi-modal (e.g. visual, auditory, haptic, etc) observations are necessary for robust execution. We present a probabilistic model, verified through physical experiments, that allows robots to acquire knowledge about the latent aspects of the workspace through language and physical interaction in an efficient manner. The model's effectiveness is demonstrated on a mobile and a stationary manipulator in real-world scenarios by following instructions under partial knowledge of object states in the environment.