Proposition of Multi-AgentPlanning using affordance extracted foundation model

Abstract

Multi-agent planning is one of the planning methods. This method allows an autonomous robot to select actions to achieve a predetermined goal in a dynamic environment. Our goal is to improve the efficiency of action selection for more dynamic environments. In this paper, we propose to extract affordances from large-scale language models and incorporate them into multi-agent planning. Since large-scale language models are trained on a large amount of text data on the web, we believe that affordances can be extracted from large-scale language models. Then, we conducted a simulation experiment in which we set the objectives to be achieved using the extracted affordances and compared the results with and without affordances. As a result, we confirmed that the use of affordances in multi-agent type planning enables us to efficiently obtain a sequence of actions to achieve the objectives.