Trust in Robot-Robot Scaffolding

Murat Kirtay, Verena V. Hafner, Minoru Asada, Erhan ÖztopIEEE Trans. Cogn. Dev. Syst 2023 · DOI

Abstract

Trust is a fundamental component of social interaction, yet formalizing it for artificial agents remains a challenge. This research proposes a model where trust is derived from internal computational effort. By treating reductions in processing costs as rewards, a robot can autonomously identify trustworthy partners. We demonstrate this model in a scaffolding scenario where a "caregiver" robot (Pepper) assists an "infant" robot (Nao) in a visual recall task, showing that trust-based intervention significantly improves learning outcomes.

Introduction

While human-robot trust is widely studied, the inverse—how robots should decide whom to trust—is less explored. In human psychology, trust is often linked to a reduction in cognitive load; we trust those who make our tasks easier and our decisions less taxing. This paper addresses the challenge of formalizing this "cognitive load" for artificial agents, demonstrating that a robot can use its own internal computational energy as a signal to assess the trustworthiness of others. Furthermore, we explore how a robot can use this trust to act as a teacher (or "scaffolder") for a novice robot, helping it navigate tasks more efficiently.

Methodology

The researchers implemented a sequential visual recall task using a High Order Hopfield Network (HOHN), an auto-associative memory system. This allows the robots to "recall" original images when presented with noisy or partial inputs.

Computational Load as Trust

"Cognitive load" was defined as the computational energy required for recall—specifically, the number of bits that changed state before the network reached a stable conclusion. Using a SARSA reinforcement learning algorithm, the robot generated internal rewards:

  • Positive Reward (+1): If an action or a partner's suggestion led to a state with a lower computational load.
  • Negative Reward (-1): If the load increased.

Scaffolding and Theory of Mind (ToM)

In the robot-robot interaction phase, a "Caregiver" robot (Pepper) utilized a Theory of Mind module—a matrix tracking the "Infant" robot's (Nao) behavior—to decide when and how to intervene. This decision-making was based on Pepper's previously learned "trust" (the optimized Q-matrix from the initial experiment).

Experimental Setup

The study utilized physical Pepper and Nao humanoid robots to ensure the model could handle real-world constraints like camera noise and varying resolutions. The task involved identifying specific icons within a complex visual environment.

The five base visual patterns used for training the robots' associative memories.

The robots were presented with a 4x5 grid containing 20 states. These states consisted of the original patterns, noisy versions, or cropped versions, requiring varying levels of computational effort to resolve.

The 4x5 visual scene grid used in the recall task, displaying various levels of noise and occlusion.

Key Results

  • Instructor Selection: In the first experiment, Pepper successfully identified the "Reliable" instructor (an oracle that always suggested low-load patterns) by achieving the highest average cumulative reward (104.2) compared to "Less-reliable" (83.0) or "Random" (69.4) instructors.
  • Scaffolding Success: When Pepper (equipped with ToM and Trust) helped Nao, the infant robot achieved a significantly higher average reward (78.0) and required fewer help requests (1100) than when working with less sophisticated caregivers.
  • Social Teaching vs. Data Transfer: Interestingly, "social teaching" (scaffolding) was more effective than simply copying a policy from one robot to another. Because Nao and Pepper have different hardware and camera noise, a direct "brain dump" was less effective than a guided learning process.

Conclusion and Future Work

By grounding trust in resource optimization, robots can develop autonomous social strategies without explicit human labeling. This approach suggests that trust is a functional mechanism for minimizing internal entropy. Future directions include scaling the model to more complex tasks, exploring trust in larger groups containing humans and robots, and moving toward a "robot-culture" where knowledge is passed down through generations via social learning.

Summary updated

BibTeX
@article{kirtay2023trust,
  author = {Murat Kirtay and Verena V. Hafner and Minoru Asada and Erhan Öztop},
  title = {{Trust in Robot-Robot Scaffolding}},
  journal = {IEEE Trans. Cogn. Dev. Syst.},
  year = {2023},
  doi = {10.1109/tcds.2023.3235974},
}
Figure