Prompting Robot Teams with Natural Language

Abstract

A framework to prompt multi-robot teams with high-level tasks expressed in natural language, using the reasoning of language models in settings where such models cannot be called at deployment time. A task is represented as a deterministic finite automaton, and recurrent neural networks are used to encode the automaton — distilling the logic and sequential decompositions of sub-tasks into a policy that supports decentralized, real-time operation on-board each robot.

Paper