Abstract
Adaptive automation aims to dynamically shift tasks between humans and machines to optimize performance and manage cognitive workload. A critical challenge in these systems is the "hand-off" of control, which often results in lost situational awareness or cognitive overload. This research investigates whether aligning machine action choices with human behavioral patterns improves team effectiveness. By developing a gamified data collection environment and a novel seven-step design framework, the study establishes a method for profiling human operators and measuring the similarity of machine-generated trajectories. Preliminary results demonstrate the feasibility of group-based profiling and provide a foundation for future experimental validation of human-machine similarity effects.
Introduction
In complex operational environments, such as managing a fleet of unmanned vehicles (UVs), human operators often rely on automation to handle secondary tasks. However, the transition of control between the human and the system—the "hand-off"—is a frequent point of failure. If an operator is excluded from the loop for too long, they lose situational awareness; if they are too involved, they face cognitive exhaustion.
The core challenge lies in Adaptive Automation: systems that dynamically shift tasks based on the current context. While existing research focuses on the timing of these transitions, this paper explores the nature of the machine's actions. Specifically, it asks whether a machine that "thinks" and acts like its human partner facilitates smoother transitions. The author highlights that the act of "handing off" is itself an inherent task that adds to a human's workload. By aligning machine actions with human expectations, the system aims to reduce the "hidden" cognitive cost of these transitions.
Methods
The author proposes a multi-stage approach to bridge the gap between human and machine behavior:
1. Function-to-Task Design Process Model
A central component of the methodology is a novel seven-step visual framework designed to identify specific points in a system where transitions between human and machine can be optimized. This model guides the development of the research environment by decomposing high-level goals into specific, manageable tasks.
2. Gamification and Data Collection
To capture high-quality human decision-making data, the author developed Space Navigator, a tablet-based game. In the game, users draw trajectories to interact with the system, providing a rich database of "human-like" actions that are more portable and accessible than standard external benchmarks.
3. Player Profiling and Similarity Metrics
The system uses k-means clustering to group users based on their playstyles, with feature selection guided by a Fisher score algorithm. To quantify how "similar" a machine's action is to a human's, the author adapted the Windowed Hausdorff Distance, allowing for the comparison of trajectory shapes and lengths even when they are not identical.
Key Results
As this research was presented as part of a Doctoral Consortium, the results focus on the framework's feasibility and initial system development:
- Model Completion: The seven-step Function-to-Task design model is fully developed and was successfully used to construct the research environment.
- Feasibility of Profiling: Preliminary results suggest that while predicting individual actions is difficult, group-based player profiling via k-means clustering is a viable path forward for tailoring automation to user tendencies.
- System Development: The Space Navigator platform is fully operational, providing a simplified way to collect complex trajectory data that was previously difficult to capture.
Discussion and Future Work
The most provocative idea presented is that "optimal" machine performance might not be the ultimate goal for human-machine teams. If a machine's optimal path is alien or confusing to a human, the cost of the human re-orienting themselves during a hand-off might outweigh the machine's technical efficiency.
Future directions include: * Integrating workload-based triggers to decide when automation should activate based on the operator's real-time mental state. * Conducting formal experiments to determine if "similar" automation increases trust and reduces perceived workload compared to dissimilar, "optimal" approaches. * Exploring how these findings apply to training—specifically, whether a machine that acts "slightly better" than the human can help bridge the gap to expert-level performance.
Summary updated