A Model Reconciliation Approach for Collaborative Recovery Behavior in Explainable Shared Control

Kurzfassung

This methodological article addresses the challenge of explaining and resolving unexpected robotic behavior in assistive robotics. Inspired by mental modeling techniques, unexpected behavior is conceptualized as a result of diverging mental models between the user and the robot due to incorrect beliefs of the robot. The explanation and recovery process is then understood as a model reconciliation within the human-robot-team. The framework is implemented using foundation models. A Large Language Model (LLM) serves as a Natural Language Interface to process user questions and provides explanations for unexpected behavior based on the robot’s knowledge. A Vision Language Model (VLM) verifies the user’s contradicting beliefs and suggests recovery behavior to update the mistaken robot’s model. The implementation accounts for false beliefs of the robot about the symbolic state of the current environment, as well as objects and their affordable actions within the environment. The method is validated against a naive VLM, which can only provide explanations based on camera images but not the robot’s modules. Accuracy scores for the proposed method lie at 100.00% for explaining failed object localization and at 100.00% for explaining wrong beliefs of the symbolic state of the world. The accuracy for correct recovery suggestion lies at 56.25%. In comparison, the accuracy for the naive VLM lies at 60.00%, 0.00% and 25.00%, respectively.