Trust Dynamics and Verbal Assurances in Human Robot Physical Collaboration

Abstract

Trust is the foundation of successful human collaboration. This has also been found to be true for human-robot collaboration, where trust has also influence on over- and under-reliance issues. Correspondingly, the study of trust in robots is usually concerned with the detection of the current level of the human collaborator trust, aiming at keeping it within certain limits to avoid undesired consequences, which is known as trust calibration. However, while there is intensive research on human-robot trust, there is a lack of knowledge about the factors that affect it in synchronous and co-located teamwork. Particularly, there is hardly any knowledge about how these factors impact the dynamics of trust during the collaboration. These factors along with trust evolvement characteristics are prerequisites for a computational model that allows robots to adapt their behavior dynamically based on the current human trust level, which in turn is needed to enable a dynamic and spontaneous cooperation. To address this, we conducted a two-phase lab experiment in a mixed-reality environment, in which thirty-two participants collaborated with a virtual CoBot on disassembling traction batteries in a recycling context. In the first phase, we explored the (dynamics of) relevant trust factors during physical human-robot collaboration. In the second phase, we investigated the impact of robot’s reliability and feedback on human trust in robots. Results manifest stronger trust dynamics while dissipating than while accumulating and highlight different relevant factors as more interactions occur. Besides, the factors that show relevance as trust accumulates differ from those appear as trust dissipates. We detected four factors while trust accumulates (perceived reliability, perceived dependability, perceived predictability, and faith) which do not appear while it dissipates. This points to an interesting conclusion that depending on the stage of the collaboration and the direction of trust evolvement, different factors might shape trust. Further, the robot’s feedback accuracy has a conditional effect on trust depending on the robot’s reliability level. It preserves human trust when a failure is expected but does not affect it when the robot works reliably. This provides a hint to designers on when assurances are necessary and when they are redundant.

Measuring and Understanding Trust Calibrations for Automated Systems: A Survey of the State-Of-The-Art and Future Directions

“…Screening of dangerous objects/subjects [4, 14, 15, 19, 36, 51, 54-56, 85, 94, 107, 117], detection of system malfunctions [89,90], crime prevention [10], recidivism prediction [132], watch a video of a house search [66,83] Transportation Responding to take-over requests [2, 6, 7, 47, 61, 68-70, 77, 84, 98], collusion avoidance [8], managing (air) traffic [31,118], pedestrians interaction with AV [48], observe AVs [64,119], drive in driving simulator [87,96,136] Military Screening tasks [17,33,44,63,81,82,127,130,135,139,140], gathering of information [58], mission planning [92], human-AI collaboration for search and destroy missions [116] Production Improving production [73,133,143], disassembly [5], moving objects [34,35,45], demand forecasting [39], harvesting [113], quality checks [142] Gaming trust game [3,23], collaboration game [24], flanker task…”

“…While the development of many complex automated technologies is driven by artificial intelligence, we cannot make specific claims about these systems. For example, some papers examined robots (see e.g., [3,5,6,22,24,50,130]). From previous work, we know that robots are increasingly equipped with artificial intelligence (see e.g., [111]).…”

“…Some authors work with correlations to get a better sense of perceived trustworthiness and system capabilities [5,12,24,39,66,67,81,94,130,139,142]. Here the larger the correlation between perceived trustworthiness and system capabilities, the better the calibration.…”

“…Moreover, Kraus and colleagues [69] found that providing prior information affected users differently depending on their need for cognition. To achieve an alignment after an interaction, some studies provide feedback about the user and system performance (see e.g., [5,14,82]).…”