Domain Concretization from Examples: Addressing Missing Domain Knowledge Via Robust Planning

The assumption of complete domain knowledge is unwarranted for robot planning and decision-making in the real world. Incompleteness in domain knowledge may come from design flaws or arise from domain ramifications or qualifications. In such cases, traditional planning methods can produce highly undesirable behaviors. Addressing the planning problem under incomplete domain knowledge is challenging since the agent has no clue about what information is missing. This is a type of unknown unknowns, which differs significantly from partial observability, a type of known unknowns. In this work, we assume that the missing information is encoded in a set of examples or teacher demonstrations. We formulate the problem of domain concretization with these examples as an inverse problem to domain abstraction. Given a domain model provided initially, when the model does not conform with the teacher demonstrations, our method searches for a candidate model set that refines the initial model under a minimalistic and deterministic model assumption. For new problems, it generates a robust plan with the maximum probability of success under the set of candidate models. Together with a standard search formulation in the model-space, we propose a heuristic-based search method and also an online version of it to reduce the search time. We evaluated our approach with several International Planning Competition (IPC) domains and a simulated robotics domain where incompleteness was introduced by removing domain features from the complete models. Results show that our methods increase the success rate of planning without significantly impacting the plan cost.