Astronauts operate in an environment with multiple hazards that can develop into life-threatening emergency situations. Managing stress in emergencies may require cognitive resources and lead to diminishing performance. Stress training aims to maintain performance under stress by methodically increasing stressor levels to build inoculation against stress. An adaptive virtual reality (VR) training system was developed with real-time stress detection by using machine learning on psychophysiological responses. Using a VR simulation of a spaceflight emergency fire, stress classifications were used to trigger adaptations of the VR environmental stressors (e.g., smoke, alarms, flashing lights), with the goal of maintaining a manageable level of stress during training. Fifty-seven healthy subjects underwent task training over eight trials with adaptive training (adaptive, n=19); results were compared to trials with predetermined gradual increases in stressors (graduated, n=18), and with trials with constant low-level stressors (skill-only, n=20). Stress responses were measured through heart rate, heart rate variability (i.e., root mean squared of successive differences (RMSSD), low frequency to high frequency (LF/HF) ratio), and task performance (distance-from-fire). Heart rate decreased and RMSSD increased pre-post training for all experimental conditions. The LF/HF ratio decreased pre-post training for the adaptive condition, but not in the other conditions. Results suggests that all conditions had lower stress, but the adaptive condition was more successful. Task performance showed a marginal increase across trials for the adaptive condition. Preliminary results suggest that training with the adaptive stress system can prepare individuals for responding to stressors better than skill-only and graduated training.