Courteous Anticipation: Improving Long-Lived Task Planning in Persistent Shared Environments

This paper studies long-horizon task planning for robots that repeatedly receive tasks in a persistent shared environment. Its courteous anticipatory planner uses a model-based planner to generate candidate plans, then scores them by combining immediate cost with estimated future cost across all robots. Independent per-robot learned estimators avoid combinatorial joint rollouts and allow modular expansion as robots are added. In two PDDL domains, the method reduced total sequence cost by 10.43% versus myopic planning and 4.03% versus selfish anticipation in a two-robot home setting, and by 17.41% and 13.24% in a three-robot restaurant setting.