Introduction
Modern complex missions, such as search and rescue or large-scale warehouse logistics, increasingly rely on Multi-Operator Multi-Autonomy (MOMA) teams. Unlike simple dyadic interactions (one human, one robot), MOMA environments involve "teams of teams" where trust is not just a direct feedback loop but a social phenomenon. In these settings, an operator's trust in a robotic agent is influenced not only by their personal interactions but also by the experiences and reports of their teammates. Existing trust models often fail to capture this propagation of trust, leading to inaccuracies in predicting team dynamics. This paper introduces the Trust Inference and Propagation (TIP) model, a mathematical framework designed to track trust evolution across multi-human multi-robot teams by accounting for both direct and indirect experiences.
The TIP Model
The TIP model represents trust as a Beta random variable, characterized by parameters $\alpha$ (representing positive experiences) and $\beta$ (representing negative experiences). This statistical approach allows for a probabilistic interpretation of trust levels. The core innovation of TIP lies in its dual-update mechanism:
- Direct Experience: When a human operator interacts directly with a robot, the model updates the $\alpha$ and $\beta$ parameters based on the robot's performance (success or failure) in the assigned task.
- Indirect Experience (Propagation): Trust is updated when an operator receives a trust report from a teammate. This propagation is governed by the "Trust Gap"—the difference between the teammate's report and the operator's current expectation—and the level of trust the operator has in that teammate.
The model parameters, including learning rates and discounting factors for social information, are optimized using gradient descent to best fit empirical data.
Experimental Evaluation
To validate the TIP model, the researchers conducted a human-subject experiment involving 30 participants paired into 15 teams. The teams performed a simulated threat detection task using two drones (Drone A and Drone B) with pre-programmed accuracies of 90% and 60%, respectively. Over 15 sessions, participants were randomly assigned a drone and required to report their trust in the drone they used (direct), the drone their partner used (indirect), and their partner. This setup provided a ground truth for how humans "learn" robot reliability through both personal use and social observation.
Results
The TIP model demonstrated superior performance compared to a baseline model that only considered direct experience. Quantitatively, TIP achieved a significantly lower Root-Mean-Square Error (RMSE) for both robotic agents:
- Drone A: RMSE of 0.057 (TIP) vs. 0.085 (Baseline).
- Drone B: RMSE of 0.082 (TIP) vs. 0.107 (Baseline).
Paired-sample t-tests confirmed these improvements were statistically significant ($p < .001$). Qualitatively, the model's 90% probability interval successfully captured the majority of trust fluctuations reported by participants, including instances of "trust oscillation" where opinions shifted rapidly.
Discussion and Future Work
While the TIP model provides a robust framework for multi-agent trust, several challenges remain. The model assumes that human agents are cooperative and truthful; in competitive or adversarial environments, deceptive reporting could undermine trust propagation. Additionally, the current model treats trust as a single-dimensional metric, whereas real-world trust is often multi-faceted (e.g., trusting a robot's navigation but not its sensors). Future research will focus on scaling the model to larger swarms and incorporating multi-dimensional trust metrics to better reflect the complexities of human-robot teaming.
Summary updated