Abstract
In this work, we present a new planning formalism called Expectation-Aware (EA) planning for decision making with humans in the loop where the human's expectations about an agent may differ from the agent's own model. We show how this formulation allows agents to not only leverage existing strategies for handling model differences like explanations and explicability, but can also exhibit novel behaviors that are generated through the combination of these different strategies. Our formulation reveals a deep connection to existing approaches in epistemic planning, specifically leveraging classical planning compilations to solve EA planning problems. We empirically show how our approach provides a computational advantage over earlier approaches that rely on search in the space of models.
Introduction
In human-robot collaboration, a major hurdle is knowledge asymmetry. A human supervisor often maintains an incorrect or outdated "mental model" ($M_H$) of the robot's capabilities ($M_R$). For instance, in a search-and-rescue mission, a commander might believe a certain hallway is clear, while the robot knows it is blocked by fire. If the robot follows its optimal path without context, the human might find the behavior confusing or sub-optimal.
Previously, researchers addressed this through two siloed strategies:
- Explicability: The robot alters its behavior to match human expectations, even if less efficient.
- Explanation: The robot follows its own optimal plan but provides a "model reconciliation" explanation to update the human’s beliefs.
Expectation-Aware (EA) planning provides a unified formal framework to balance these strategies, enabling robots to generate "self-explaining plans."
Methodology: Compilation to Classical Planning
The core innovation of this work is a compilation-based approach that transforms the EA planning problem into a standard classical planning task with conditional effects. This avoids the computationally expensive search through the space of possible models.
The Augmented Model
The state space is augmented to include both physical fluents (world state) and "meta-fluents" representing the human's beliefs about the robot's model. This allows the planner to track how actions affect both the environment and the human's mental state.
Epistemic Side Effects and Communication
The framework recognizes that physical actions can have communicative side effects. For example, if a robot opens a door the human thought was locked, the human's model is updated implicitly. Additionally, the model treats explicit communication (e.g., "I am carrying two boxes") as actions with specific costs and effects. By treating communication as just another action, the planner can perform a cost-benefit analysis to decide whether to act predictably or explain its true capabilities.
Experimental Evaluation
The framework was evaluated using modified benchmarks from the International Planning Competition (IPC), including Blocksworld, Elevator, Gripper, Driverlog, and Satellite. Problem instances were generated by introducing 5 to 10 random differences between the robot's model and the human's assumed model.
As shown in the table above, the compilation approach significantly outperforms previous "model space search" methods. The results demonstrate that the compilation method achieved higher coverage and drastically lower runtimes across all tested domains. In the Elevator domain, for instance, it solved 15/15 problems in an average of 59.20 seconds, compared to the search method's 1/15 coverage and 3382.46 second runtime.
Conclusion and Future Work
Expectation-Aware planning successfully unifies explanation and explicability into a single, computationally efficient framework. By compiling these problems into classical planning, we leverage existing high-performance solvers to handle complex human-aware scenarios. Future work will focus on relaxing deterministic assumptions, handling nested beliefs, and incorporating model uncertainty where the robot may not perfectly know the human's mental model.
Summary updated