Dialogue-driven Multi-Agent Activity Planning

Abstract

A fundamental challenge in robotics is to build a general-purpose system with multiple agents that can perform a wide range of tasks based on specifications provided in natural language. This work presents a novel dialogue-driven activity planning framework for multiagent scenarios. We present a method that accepts commands from a user in natural language and translates it to an intermediate form called a state plan by leveraging large language models. We further experiment with chain-of-thought prompting to improve the translation from natural language to state plans. In conjunction with an action model, this state plan is utilized by a constraint-based generative planner called ctBurton which outputs a full grounded plan in the form of a state and control trajectory. We demonstrate the utility of our method across three different scenarios– a presentation system, search-and-rescue, and multi-agent assembly– along with experiments on its scalability.