We are at a methodological hinge. Military planners and roboticists now speak in the same breath about hybrid tactics — coordinated plans that treat humans and autonomous systems as coequal members of a team rather than as tools under human command. To evaluate those tactics before they are fielded requires more than larger wargames and faster physics engines. It requires simulations that faithfully represent human decision processes, the failure modes of autonomy, and the sociotechnical frictions that emerge when the two mix.
Two broad purposes drive hybrid tactics simulations. The first is operational: to discover tactics, techniques and procedures that exploit complementarity between humans and machines under realistic constraints. The second is cognitive: to study trust, workload, and shared mental models so that doctrine, interfaces and training can be adapted to the limits of real people and real autonomy. Treating these as separate objectives is a mistake. Tactical performance and human factors are coupled. If a simulation overestimates an autonomy stack or underestimates human cognitive delay it will produce tactics that fail in the field.
Recent activities in both research and defense experimentation illustrate these objectives. DARPA has explicitly opened a line of work to produce digital twins of human-AI teams with the goal of generating and evaluating realistic human teammate simulacra. That program envisions using generative AI to create diverse, naturalistic human behavior models that can be paired with automated agents to stress-test teaming behaviors in proxy operational settings.
At the academic end, reviews of the literature emphasize the breadth of human roles in military teams and the corresponding need for simulations that span from high-level command decision making down to dismounted, in-situ operators. The necessary research agenda runs from basic human factors and interface design through to scalable simulation platforms that can host many heterogeneous agents while preserving fidelity of human behavior. Simplistic supervisor-only models will not suffice.
Operational experiments conducted by services provide a complementary lesson. Recent multi-domain exercises have embedded swarms, unmanned ground vehicles and autonomous sensors into large scale experiments in order to evaluate human-machine integration in realistic mission flows. Those events show how tactical ideas emerge when autonomy is present at scale, but they also expose a persistent gap: the laboratory or exercise can show a capability works when closely supervised by its developer or a specialist operator, yet fail when handed to generalist warfighters with different mental models and different tolerances for risk. Simulations therefore must intentionally model heterogeneity in operator skill and behavior.
What should a credible human-robot hybrid tactics simulation include? I propose five minimum design requirements.
1) Human behavioral diversity. Simulated humans should not all be Bayesian optimizers or identical heuristics. They must display bounded rationality, variable reaction times, communication delays and occasional systematic biases. Recent modeling work shows the importance of representing suboptimal human behavior explicitly and allowing autonomy to adapt online to such suboptimality rather than to an abstract idealized teammate. Incorporating this produces more robust mixed-initiative tactics.
2) Trust and calibration metrics embedded in the loop. Trust is not a cosmetic variable. It changes what tasks humans delegate, how often they intervene and how tightly they monitor autonomy. Formal trust-inference models for multi-human multi-robot teams can be implemented inside a simulation to measure how trust propagates across a formation and how it affects mission outcomes. These metrics should be logged alongside classical performance measures.
3) Generative human digital twins for scale. To explore many tactical permutations you need many plausible humans. Programs that propose generative AI for producing realistic teammate simulacra are promising because they allow scaling experiments without exhaustive live-subject testing. But generative models must be validated against empirical human subject data and must include behavioral noise and adversarial responses to avoid overfitting tactics to a synthetic population.
4) Human-in-the-loop embodiment when it matters. When interaction modalities matter, for instance in dismounted small unit tactics or in joint human-robot manipulation tasks, immersive human-in-the-loop platforms provide better fidelity. Recent simulation platforms that couple VR embodiment to robotic control pipelines demonstrate that human responses in embodied scenarios can differ materially from desktop command interfaces. Use embodiment selectively where outcomes are sensitive to sensory-motor coupling.
5) Red teaming and failure-mode injection. The most valuable simulations are those that intentionally break assumptions. Inject sensor spoofing, communication denial, degraded autonomy and deceptive social signals. Evaluate how both human judgment and autonomy adapt, and whether the tactics depend on brittle assumptions about perfect communications or reliable sensing.
Even when these elements are present, there remain critical caveats. First, simulation fidelity is necessary but not sufficient. There is a persistent risk of the credibility gap where simulated performance becomes a marketing milestone rather than an operational guarantee. Experimental design must therefore include validation phases that compare simulation predictions to instrumented live trials with representative users.
Second, simulations can normalize certain ethical and legal assumptions unintentionally. Automated delegation rules, escalation pathways and lethal decision criteria encoded for expedience in a simulation can ossify into doctrine if the downstream organizational checks are not enforced. Simulation designers therefore share a moral responsibility. Simulations should not only measure effectiveness but also include observability for accountability. They should log decision provenance, record human overrides, and permit post hoc audit of autonomy behavior.
Third, there is an epistemic danger in relying on generative human twins. Generative models learn patterns in data. They can reproduce cultural or demographic biases present in training sets, and they can fail spectacularly when faced with novel stressors outside their training distribution. Treat generative agents as hypothesis generators, not as final verdicts on human behavior.
Finally, there is a strategic dimension. Simulations are a laboratory for tactics but they are also a political artifact. How we simulate adversaries, civilian populations and the fog of war reflects normative choices. Transparent documentation of scenario assumptions, public release of non-sensitive benchmark tasks and third-party replication where possible will go a long way to preserving professional legitimacy.
To conclude, hybrid tactics simulations are indispensable for the next generation of human-robot operations. But they must be built with humility and methodological rigor. Mix generative digital twins with human-in-the-loop trials. Measure trust and adaptation, not just kill chains. Inject failure and auditability. If we treat simulations as a form of thinking tool rather than as a final test, they will help us design tactics that are not only faster and more lethal, but also safer, more accountable and more resilient when real people meet real machines in the complexity of combat.