Introduction
Effective human-robot collaboration hinges on trust. When a human teammate lacks trust in a robot, they may underutilize its capabilities, ignore its outputs, or dedicate excessive cognitive resources to monitoring it. Conversely, over-trust can lead to safety risks. A significant challenge in Human-Robot Interaction (HRI) is that trust is subjective, dynamic, and often opaque to the robot. Most research focuses on how robots should trust humans; this paper explores the inverse: how a robot can infer its own trustworthiness from a human's perspective and adapt its behavior to align with their preferences.
Methodology
The proposed framework enables a robot to adapt its behavior parameters (e.g., speed, safety margins) based on an "inverse trust" metric derived from observable interactions.
Inverse Trust Estimation
The robot maintains a running trust score based on task success and human interventions. Successful task completion increases the score, while failures or interruptions by the human decrease it. Two thresholds classify the robot's current behavior as trustworthy, untrustworthy, or unknown.
Behavior Adaptation via Case-Based Reasoning (CBR)
When a behavior is flagged as untrustworthy, the robot must find a new configuration. Rather than exhaustive trial and error, the system employs Case-Based Reasoning. The robot stores "cases" consisting of a problem (a sequence of failed behaviors) and a solution (the behavior that eventually satisfied the operator).
To identify the most relevant past experience, the robot compares its current sequence of failed attempts—the "search path"—to those stored in its memory. This comparison uses a similarity metric for sets of evaluated behaviors to identify similar operator profiles.
By identifying similar search paths, the robot can effectively "fingerprint" the current operator's preferences and jump directly to a behavior likely to be trusted, bypassing redundant exploration.
Experimental Evaluation
The framework was evaluated in the eBotworks simulation environment across two scenarios:
- Movement: Navigating to waypoints with adjustable speed and obstacle padding.
- Patrol: Identifying objects with adjustable speed, padding, scan time, and scan distance.
Simulated operators were programmed with distinct preferences (e.g., "speed-focused" vs. "safety-focused"). The robot's goal was to minimize the number of behavior evaluations required to find a trustworthy configuration.
Results
The CBR-based adaptation was compared against a "Random Walk" baseline. The results demonstrate significant improvements in adaptation efficiency across all scenarios and operator types.
In the Movement scenario, CBR reduced the average number of evaluations from 20.3 to 1.6. In the more complex Patrol scenario, the improvement was even more pronounced, reducing evaluations from 344.5 to 9.9. These results indicate that the "search path" is a highly effective proxy for identifying human preferences.
Discussion and Future Work
The use of interruptions as a high-signal negative reward allows the robot to learn without explicit trust scores. However, the current model assumes all interruptions are performance-related, which may not hold in dynamic missions where goals change. Furthermore, the system faces a "cold start" problem, requiring initial random exploration to populate the case base. Future work will explore incorporating occasional explicit feedback and reasoning about team goals to better distinguish between behavior-driven and environment-driven trust fluctuations.
Summary updated