Exploring the integration of large language models in human-robot collaboration: Effects on performance, mental stress, and trust

As large language models (LLMs) become increasingly integrated into robotic systems, understanding their influence on human-robot collaboration (HRC) is critical for designing effective and user-centered human-robot interactions. This study investigates the impact of LLM-enhanced robotic systems on users' performance, mental stress, and trust during collaborative tasks. Participants engaged in two representative HRC scenarios, including object delivery and instruction following, under two experimental conditions: with and without LLM support. Performance was measured through task completion time and number of verbal commands; mental stress was assessed using both subjective (NASA-TLX) and objective (galvanic skin response, GSR) measures; and trust was evaluated through the SHAPE Trust Index and eye-tracking metrics (blink rate and duration). Results showed that LLM integration significantly improved task efficiency and reduced subjective mental stress, particularly mental demand, effort, and frustration. Participants also reported higher levels of trust in the LLM condition across dimensions such as usefulness, reliability, accuracy, and ease of use. Interestingly, GSR data indicated elevated physiological arousal, possibly suggesting increased engagement or positive emotional activation, while eye-tracking measures showed no significant differences. These findings highlight the potential of LLMs to enhance HRC by enabling more natural communication, reducing mental workload, and increasing user trust, while also pointing to the need for improved system transparency to support deeper understanding and sustained trust.