Verifiable Grounding and Execution of Natural Language Instructions

Adrian Boteanu


Robots are increasingly often expected to work along with humans. Natural language enables bi-directional interaction: for users to specify tasks and for the system to provide feedback. A significant challenge particular to this situated interaction is establishing correspondence between language and their physical meaning such as actions and objects, known as grounding. As both tasks and environments increase in complexity, the potential for ambiguity in interpreting the user’s statements increases.

I will present a grounding model which combines both physical and Linear Temporal Logic (LTL) representations to ground instructions. It allows for a formal specification to be generated from the grounding process. This specification is synthesized into a controller guaranteed to accomplish the task. Conversely, if synthesis is unsuccessful, it reveals problems such as logical inconsistencies in the specification or discrepancies between the specification and the physical environment.

In this latter case, the robot conveys these issues through natural language by referencing the physical environment and incorporates the user’s responses back into the specification. This robot-driven interaction enables the user to iteratively correct the grounded specification without requiring knowledge of the underlying representation.