Authors:
Polyana Costa
1
;
Pedro Santos
1
;
José Boaro
1
;
Daniel Moraes
1
;
Júlio Duarte
2
and
Sergio Colcher
1
Affiliations:
1
Pontifícia Universidade Católica do Rio de Janeiro, Rio de Janeiro, Brazil
;
2
Instituto Militar de Engenharia, Rio de Janeiro, Brazil
Keyword(s):
Ad Hoc Teamwork, Large Language Models, Prompt Engineering, Autonomous Agents.
Abstract:
Ad Hoc Teamwork environments are dynamic spaces where agents engage in activities, make decisions, and collaborate with teammates without prior coordination or complete knowledge of tasks. To effectively operate in such an environment, an ad hoc agent must be equipped with robust reasoning and planning mechanisms. Since Large Language Models (LLMs) are known for their generalization abilities, this study showcases their application in ad hoc scenarios. By modeling the robot’s actions using LangChain Tools, building a semantic map, and capturing human communication interactions, we tested the LLM reasoning capabilities in three simulated scenarios involving humans and a robot. In each case, after providing contextual information, we build a meta-prompt with the question: ‘How can the Robot help?’. By conducting these tests, this study highlighted the LLM’s ability to infer tasks and craft action plans even in the absence of explicit verbal commands.