Model Reconciliation through Explainability and Collaborative Recovery in Assistive Robotics

Technical Summary: Model Reconciliation through Explainability and Collaborative Recovery in Assistive Robotics

Problem Statement
In assistive robotics and shared control scenarios, unexpected robot behavior often stems from a divergence between the human user's mental model of the world and the robot's internal model. When a robot fails to execute a task (e.g., grasping an object), the user may not understand the robot's reasoning, leading to a breakdown in trust and collaboration. Existing model reconciliation approaches often assume the robot possesses a complete, formal model of the human's mental state, which is unrealistic in dynamic assistive environments. Furthermore, while explanations can clarify why a failure occurred, they do not inherently provide a mechanism for the human to correct the robot's erroneous beliefs to recover from the failure.

Methodology
The authors propose a model reconciliation framework that leverages Foundation Models (specifically Large Language Models [LLMs] and Vision-Language Models [VLMs]) to bridge the gap between human and robot mental models without requiring a pre-defined formal model of the user. The framework operates in three stages:

1. Start & Explanation Generation:

   • The process is initiated by a natural language query from the user (e.g., "Why can you not grasp the greenish cup?").
   • An LLM workflow extracts semantic entities (objects, actions, adjectives) from the query.
   • The system performs a four-step search against the robot's internal models:
     • Object Database Check: Verifies if the object exists in the robot's static knowledge base.
     • Action Model Check (Pass 1): Determines if the action is theoretically possible given the current world state.
     • World Model Check: Verifies if the specific object instance is currently perceived and localized by the robot.
     • Action Model Check (Pass 2): Identifies if the action is blocked by unmet preconditions (e.g., a gripper not being free) or if the action is simply not trained.
   • Based on this search, the system predicts the specific type of model divergence (e.g., the robot doesn't know the object exists vs. the robot sees the object but cannot act) and generates a natural language explanation.

2. Collaborative Recovery:

   • If the explanation reveals a discrepancy the user can correct (e.g., the robot missed an object or has a wrong state belief), the user provides a rebuttal (e.g., "But the drawer is closed!").
   • VLM Arbitration: A Vision-Language Model acts as an "oracle" to validate the user's claim against the robot's camera feed.
     • For Perception Errors (DSO): If the user claims an object is present, the VLM attempts to match it in the image. If successful, it suggests robot movements (e.g., moving the wheelchair) to improve the view for detection. If the VLM cannot confirm but the user insists, the system may overwrite the world model to include the object.
     • For State Errors (DSA): If the user corrects a state (e.g., "The microwave is open"), the system maps this natural language correction to the symbolic world model (PDDL) and updates the action graph accordingly.

3. Implementation:

   • The framework was implemented on a wheelchair-based mobile manipulator designed for people with motor impairments.
   • The system utilizes a local open-weights VLM (Mistral-Small-3.2-24B-Instruct) with Chain of Thought and Few-Shot Prompting techniques to handle semantic grounding and reasoning.

Key Contributions
The paper presents contributions across three levels:

1. Theoretical: A model reconciliation framework for shared control in assistive robotics that encompasses bidirectional updating: the robot updates the human via explanation, and the human updates the robot via correction.
2. Methodological: The integration of foundation model workflows augmented with retrieved robot module data (object databases, world models, action graphs). This approach eliminates the need for the robot to explicitly model the user's entire mental state while still predicting differences in knowledge bases.
3. Experimental: A full implementation and evaluation involving a real assistive robot and its digital twin, demonstrating the system's ability to handle daily living scenarios.

Results
The authors conducted a pilot study on a real robot and two experiments (B1 and B2) on a digital twin using a dataset of 40 episodes (120 total output conversations).

• Accuracy: The framework achieved between 78% and 100% accuracy across different unit tests. Specifically, it reached 100% accuracy in explaining object localization failures, 78.79% in communicating symbolic state errors, and 78.12% in generating recovery suggestions.
• Comparison: In Experiment B1, the proposed method significantly outperformed a "naive" VLM baseline that lacked access to the robot's internal models, demonstrating that the LLM's ability to retrieve and synthesize specific robot module data is crucial for accurate explanations.
• Semantic Grounding: Experiment B2 showed that providing a translation dictionary for uncommon robot-specific terms (e.g., mapping "ikea bagganas" to "cabinet") increased accuracy in action reconciliation from 78.8% to 92.4%.
• Limitations: The authors note that recovery movement suggestions (spatial reasoning) showed lower accuracy, likely due to limitations of the specific VLM used. Additionally, adjectives in queries sometimes caused confusion, and the study was limited to a specific set of daily living scenarios.

Significance
The paper claims that this work provides a practical workflow for explaining unexpected robot behavior and enabling collaborative recovery in assistive robotics. By using foundation models to predict and explain model divergences, the system allows humans to understand the robot's beliefs and actively correct them, thereby restoring shared understanding without requiring complex, pre-defined user modeling. The authors emphasize that while the current implementation is modest and limited to specific scenarios, it establishes a viable pipeline for transparency and trust in human-robot teams, paving the way for future work on plan optimality, goal misalignment, and more diverse object instances.