Human–Autonomy Teaming: A Review and Analysis of the Empirical Literature

TL;DR

This integrative review synthesizes 76 empirical articles on human–autonomy teaming (HAT), defines HATs as interdependent human(s) and autonomous agent(s) occupying distinct team roles, and maps findings using an I-M-O framework. It finds generally positive effects of higher agent autonomy, reliability, interdependence, and training, but human–human teams usually outperform HATs and transparency/communication quantity show mixed effects.

Key findings

• The review identified 76 empirical articles; all were simulation/lab studies, none field-based. Of these, 57 were military- or emergency-related simulations, 17 were nonmilitary/emergency simulations, and 2 could not be classified; the most represented testbeds were CERTT-UAS-STE (18 articles) and MIX (15 articles).
• Agent autonomy was coded as high in 26 articles, partial in 31, and multiple levels in 15. Higher agent autonomy generally had positive effects (e.g., easier to work with, higher communication efficiency, stronger coordination, higher performance, lower workload), though some studies found a moderate level ideal; reliability was consistently positive, while transparency had mixed effects (e.g., improved performance/trust/understanding but also complacency, conflict, and workload).
• Human–human teams generally outperformed HATs; interdependence led to positive outcomes; higher task difficulty led to negative outcomes such as higher workload and lower performance; training produced consistently positive results; communication quality was associated with stronger team functioning, but communication quantity findings were mixed.
• Dependent variables studied most often were team and individual performance (70 studies), workload (39), trust (24), situation awareness (23), team coordination (15), and shared mental models (6); little research addressed Hackman’s team viability and individual learning/growth outcomes.

Methods

• Narrative/integrative literature review with four-step search. Inclusion required empirical research, sufficient description to code agent autonomy, and at least one autonomous agent and one human working interdependently toward a common goal; the agent had to reach at least Level 5 on Parasuraman et al.’s (2000) automation continuum.
• Searches covered ACM Digital Library, IEEE Xplore, and Sage Journals: initial search returned 1,942 articles, reduced to 476 after full-text “team” filtering, yielding 33 included; author-focused search returned 930 potential articles, yielding 31 included; Google Scholar search returned 152 articles, yielding 8 included; adding “human–automation interaction” yielded 4 more; total unique articles = 76.
• Coding was inductive and iterative, organized within the I-M-O model of team effectiveness. After calibration, 30% of remaining articles were randomly selected for independent coding; inter-coder agreement was 89% overall.

Limitations

• Stated: the search may have missed relevant articles; using “teamwork” as a search term may have constrained coverage; a disproportionate number of papers came from two laboratories/testbeds (CERTT-UAS-STE 24%, MIX 20%); it is unclear whether humans perceived the autonomous agents as team members; and HRI and autonomous vehicle literatures were omitted.
• Evident from the text: all included studies were simulation/lab-based, most involved short-term “zero-history” teams, often single human–single agent dyads, limited communication, and mostly partial rather than high autonomy, limiting generalizability.
• Evident from the text: no studies formally reported mediation analyses, so mechanisms linking team inputs to outputs remain largely untested.