[Paper Review] Building Second-Order Mental Models for Human-Robot Interaction

Connor Brooks, Daniel Szafır|arXiv (Cornell University)|Jan 1, 2019

Social Robot Interaction and HRIPsychology51 references13 citations

TL;DR

This paper proposes a framework for robots to infer human second-order mental models—beliefs about the robot’s behavior—by observing low-level human actions in a grid-world interaction. Using Bayesian inference on action choices, the method successfully leaks information about human mental models, demonstrating feasibility despite noisy classification, and suggests potential for improving human-robot interaction awareness through action-based inference.

ABSTRACT

The mental models that humans form of other agents---encapsulating human beliefs about agent goals, intentions, capabilities, and more---create an underlying basis for interaction. These mental models have the potential to affect both the human's decision making during the interaction and the human's subjective assessment of the interaction. In this paper, we surveyed existing methods for modeling how humans view robots, then identified a potential method for improving these estimates through inferring a human's model of a robot agent directly from their actions. Then, we conducted an online study to collect data in a grid-world environment involving humans moving an avatar past a virtual agent. Through our analysis, we demonstrated that participants' action choices leaked information about their mental models of a virtual agent. We conclude by discussing the implications of these findings and the potential for such a method to improve human-robot interactions.

Motivation & Objective

• To investigate whether low-level human actions during human-robot interactions can reveal information about their mental models of the robot.
• To develop a framework for robots to perform Bayesian inference on second-order mental models using observed human actions.
• To evaluate whether such inferred models improve goal estimation or interaction awareness in human-robot interactions.
• To explore the feasibility of closing the loop in open-loop human mental model estimation by incorporating human actions.

Proposed method

• The authors design a grid-world environment where participants navigate an avatar past a virtual robot agent with predefined behaviors.
• They define four distinct agent models (Stationary, Fixed-Goal, Chasing, Random) to represent different human mental models of the robot.
• A Bayesian inference framework is applied to update beliefs about the human’s model of the robot based on observed action sequences.
• The method uses a probabilistic model to infer the human’s underlying mental model from discrete action choices, treating actions as independent observations.
• The framework is evaluated using an online study with 40 participants, where action sequences are used to classify the participant’s true mental model of the agent.
• The approach combines action-based inference with a joint model for goal and mental model estimation to assess its impact on goal inference accuracy.

Experimental results

Research questions

• RQ1 Can low-level human action choices during human-agent interactions leak information about the human’s mental model of the agent?
• RQ2 To what extent can a robot infer a human’s second-order mental model (i.e., belief about the agent’s behavior) from observed actions alone?
• RQ3 Does incorporating inferred second-order mental models improve the accuracy of goal inference in human-robot interactions?
• RQ4 Are there systematic patterns in action-based inference errors that reflect higher-level cognitive distinctions between mental models?

Key findings

• The I-POMDP state inference method significantly outperformed chance in classifying participants’ ground-truth mental models of the agent, demonstrating that action choices do leak information about human mental models.
• Classification accuracy was approximately 50%, indicating that while information is present, inference from actions remains noisy and challenging.
• Confusion matrix analysis revealed that errors clustered between 'safe' models (Stationary, Fixed-Goal) and 'risky' models (Random, Chasing), suggesting that actions may encode abstract safety perceptions.
• The method showed marginal improvement in goal inference accuracy compared to treating the agent as a fixed object, but the difference was not statistically significant.
• The results support the feasibility of using action-based inference for second-order mental modeling, even in brief, passive interactions.
• The study provides evidence that information about human perceptions of robots is embedded in individual action selections, independent of the agent’s actual behavior.