Incorporating Mental Simulation for a More Effective Robotic Teammate

William G. Kennedy, Magdalena D. Bugajska, William Adams, Alan C. Schultz, J. Gregory TraftonAAAI 2008

Introduction

Effective teamwork requires more than just individual competence; it necessitates a "shared mental model." In human teams, members often predict what their teammates will do based on an understanding of their capabilities and the current situation. This cognitive ability, often referred to as "Theory of Mind," allows for seamless coordination without constant verbal communication.

Robots, however, typically struggle with this level of social cognition. Without the ability to predict human behavior, robotic teammates often act reactively or require explicit, time-consuming communication to stay coordinated. Researchers from the Naval Research Laboratory have addressed this challenge by developing a system that allows an autonomous robot to "think" like its human partner to improve efficiency in high-stakes environments, such as warehouse security.

Conceptual diagram of the coordination problem: a human and robot must decide which station (1 or 2) to occupy without explicit communication.

The ACT-R/E Cognitive Architecture

The researchers utilized ACT-R/E (the "E" stands for Embodied), an extension of the well-established ACT-R cognitive architecture. ACT-R is a hybrid symbolic/sub-symbolic system designed to model human cognition through a series of specialized modules.

Key technical components of the ACT-R/E system include:

  • Cognitive Modules: The system uses specialized modules for spatial reasoning, navigation ("moval"), vision, and audition.
  • Cognitive Plausibility: To ensure the robot behaves in a way a human would expect, its sensors were intentionally limited. For example, it utilizes a 140-degree field of view to match human vision, ensuring it only "knows" what a human in its position would know.
  • Procedural Module: This acts as the central hub, matching the state of various buffers to execute specific actions or "productions."
The ACT-R/E architecture, showing the interaction between cognitive modules, buffers, and the central procedural module.

Methods: The "Self as Model" Heuristic

The core innovation of this research is the implementation of Mental Simulation. To predict a teammate's choice, the robot spawns a second, internal instance of its own cognitive model. It feeds this "mental model" the teammate’s last known position and runs its own decision-making logic.

Essentially, the robot operates on the premise that humans use themselves as an initial model for others. By asking, "If I were in my teammate's shoes, what would I do?", the robot generates a "good enough" prediction of human behavior without needing a complex, separate database of human psychological profiles. This allows for proactive coordination: if the robot predicts the human will head to Station 1, it can immediately commit to Station 2, avoiding a future conflict.

Experimental Evaluation

The system was implemented on an iRobot B21r, a human-scale mobile platform equipped with an omni-directional camera and a monitor displaying an animated face to indicate its focus of attention. The approach was evaluated using two primary methods:

  1. Custom Laboratory Scenario: A warehouse security task where a human and robot patrol an area. When an alarm sounds, they must "man" two separate, distant guard stations as quickly as possible.
  2. Desktop Simulations: 25 simulated runs were used to gather statistically significant quantitative data comparing "self-centered" strategies (no simulation) vs. "collaborative" strategies (with simulation).

Key Results

The researchers compared an Independent case (where both agents simply head to the closest station) against a Collaborative case (where the robot utilizes mental simulation).

Quantitative Findings: * In 25 simulated runs, the robot using mental simulation took 3.28 fewer steps on average to complete the task compared to the independent version. * This improvement was statistically significant ($t(27.7) = 8.1492, p < .001$). * Physical robot trials confirmed these results; the simulation-enabled robot successfully predicted human intent, avoided pathing conflicts, and completed the security task faster than the reactive model.

Discussion and Future Work

While the study intentionally disallowed communication to isolate the power of simulation, real-world applications would likely benefit from a hybrid approach. Current limitations include the assumption that the human is always "self-centered"; the model does not yet account for recursive simulation (a human trying to simulate the robot).

Future directions for this research include implementing more sophisticated selection strategies, such as a cognitive version of the "minimax" algorithm, to further optimize total team travel distance and efficiency in complex, multi-agent environments.

Summary updated

BibTeX
@inproceedings{kennedy2008incorporating,
  author = {William G. Kennedy and Magdalena D. Bugajska and William Adams and Alan C. Schultz and J. Gregory Trafton},
  title = {{Incorporating Mental Simulation for a More Effective Robotic Teammate}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure