Abstract: Generating explanations to justify behavior is an essential capability for a robotic teammate. Explanations help human partners better understand the situation and maintain trust. Prior work on robot explanation generation focuses on providing the reasoning behind decision-making from the explainer's perspective, often ignoring the cognitive requirements of the explainee. In this work, we address the "progressiveness" of explanations. A progressive explanation improves understanding by limiting the cognitive effort required at each step. We present a general formulation of Progressive Explanation Generation (PEG) and provide algorithms based on alternative quantifications of cognitive effort, evaluated in a standard planning competition domain.
The Challenge of Model Reconciliation
In human-robot teaming, trust depends on the human understanding why a robot is performing a specific action. Most current Explainable AI (XAI) research focuses on "Model Reconciliation"—identifying the specific differences between the robot’s internal model ($M^R$) and the human’s mental model ($M^H$) and providing the necessary information to resolve the discrepancy.
However, providing the correct information is only half the battle. If a robot delivers a massive, complex explanation all at once, or presents facts in a jarring sequence, the human's cognitive effort spikes. This paper addresses this "dissonance" by treating explanation as a sequential process where the goal is to make the transition from the human's initial understanding to the robot's correct understanding as seamless as possible.
Defining Progressiveness in Explanations
The authors formalize PEG as a search problem in "model space." Instead of just finding the shortest set of facts to explain a plan, they look for an ordered sequence of unit changes that minimizes a "model-plan distance metric." This metric serves as a proxy for cognitive effort.
Key technical components include:
- Distance Metrics: They propose several ways to quantify effort, such as Plan Cost Difference (how much the perceived cost of the optimal plan changes) and Plan Editing Distance (how many steps in the actual plan change).
- A* Search with Heuristics: Specialized admissible heuristics allow the robot to quickly find the "smoothest" explanation path.
As shown in the figure above, a non-progressive explanation might keep the human in the dark for several steps before delivering a "shocker" that radically changes their understanding of the plan cost. A progressive explanation "nudges" the mental model gradually, leading to a smoother cognitive experience.
Experimental Evaluation
The approach was evaluated using the Rover domain, a standard benchmark from the International Planning Competition (IPC). The authors simulated human-robot teams where the "human" model was missing various features (preconditions, effects, or costs) compared to the robot's model.
Key Results
- Lower Cumulative Cognitive Load: PEG explanations consistently resulted in much lower cumulative distance values compared to standard concise explanations. This suggests that the sequence of information significantly impacts how easily a human can process updates.
- Efficiency: Despite the added complexity of sequencing, the A* search with the proposed heuristics found solutions in reasonable timeframes (often under a minute for complex rover problems), making it potentially viable for real-time interaction.
- Minimal Overhead: PEG explanations were often nearly as short as the most concise explanations, proving that you can be "smooth" without being unnecessarily wordy.
Conclusion and Future Work
This work highlights that for explanations to be effective, they must be easy to follow, not just factually correct. By prioritizing "smoothness" over simple conciseness, robots can better support their human teammates' cognitive processes.
The authors acknowledge that the current "cognitive effort" metrics are mathematical proxies. Future work will involve user studies with actual humans to validate these metrics and expanding the framework to handle interactive dialogues where the robot can adjust its explanation based on real-time human feedback.
Summary updated