Safe Difficulty: Multi-Objective Adaptive Control of Human Risk and Training Fidelity in Industrial Virtual Reality

Industrial virtual reality can expose trainees to hazardous situations without exposing them to physical danger, yet an adaptive system faces a less obvious failure mode: making training so comfortable that the difficulty to be mastered disappears. This paper introduces Safe Difficulty, a control formulation implemented in VR Guardian that jointly represents human-state risk, operational/task risk, intervention disruption, and a learning-loss cost associated with over-assistance. At each decision step, the controller selects the lowest-cost admissible intervention from scene stabilization, rotation reduction, dynamic field-of-view restriction, visual simplification, scenario slowing, micro-pauses, locomotion switching, and emergency stopping. A reproducible synthetic benchmark evaluates four policies across six industrial scenarios using 240 matched base sessions, 960 policy-session evaluations, and 115,200 policy-step evaluations. VR Guardian reduced mean modeled joint risk by 9.46% relative to Fixed Training while using no modeled session stops. Safety-First achieved the lowest risk, but VR Guardian reduced disruption by 80.0% and modeled learning-loss cost by 90.0% relative to that conservative policy. Across 27 weight settings and a ±30% action-effect stress test, the qualitative trade-off persisted. Removing the learning-loss term lowered modeled risk but increased modeled learning-loss cost by 260.0%, exposing the central protection–fidelity tension. The evidence is computational and synthetic; human efficacy remains to be established.