Model Reconciliation through Explainability and Collaborative Recovery in Assistive Robotics

TL;DR

The paper addresses the challenge of explainable human–robot collaboration in assistive shared-control by introducing a model reconciliation framework that uses an LLM to predict and explain differences between human and robot mental models without requiring a complete human model. It couples this with a recovery mechanism via a Vision-Language Model to allow the human to correct the robot’s beliefs and update its world or symbolic models accordingly. The approach is instantiated on a wheelchair-based mobile manipulator with a digital twin, integrating object databases, world models, and an action graph, and evaluated through real-robot pilots and simulation studies. Results show high accuracy in explanations and recovery, particularly when semantic representations are enhanced with a dictionary of uncommon terms, while also acknowledging limitations in perception grounding and generalizability to broader scenarios and plans.

Abstract

Whenever humans and robots work together, it is essential that unexpected robot behavior can be explained to the user. Especially in applications such as shared control the user and the robot must share the same model of the objects in the world, and the actions that can be performed on these objects. In this paper, we achieve this with a so-called model reconciliation framework. We leverage a Large Language Model to predict and explain the difference between the robot's and the human's mental models, without the need of a formal mental model of the user. Furthermore, our framework aims to solve the model divergence after the explanation by allowing the human to correct the robot. We provide an implementation in an assistive robotics domain, where we conduct a set of experiments with a real wheelchair-based mobile manipulator and its digital twin.

Figures (13)

• Figure 1: Overview of the model reconciliation framework including the bidirectional updating process between human and robot, using foundation models.
• Figure 2: Evolution of the mental models of human and robot during our model reconciliation framework. Yellow boxes represent the user's model. Grey boxes represent the robot's models. See the text for details.
• Figure 3: Overview of the explanation generation pipeline. We note that the text in the right column refers to the explanation content and not the final sentence uttered by the LLM. See Experiments and attached video for end-to-end examples.
• Figure 4: Symbolic state dictionaries for a world model with (A) a green mug and a closed microwave; (B) a green mug and an open microwave.
• Figure 5: Example prompt for matching an object with the robot database.