Objective: To contribute to the validation of virtual reality (VR) as a tool for analyzing pedestrian behavior, we compared two types of high-fidelity pedestrian simulators to a test track. Background: While VR has become a popular tool in pedestrian research, it is uncertain to what extent simulator studies evoke the same behavior as nonvirtual environments. Method: An identical experimental procedure was replicated in a CAVE automatic virtual environment (CAVE), a head-mounted display (HMD), and on a test track. In each group, 30 participants were instructed to step forward whenever they felt the gap between two approaching vehicles was adequate for crossing. Results: Our analyses revealed distinct effects for the three environments. Overall acceptance was highest on the test track. In both simulators, crossings were initiated later, but a relationship between gap size and crossing initiation was apparent only in the CAVE. In contrast to the test track, vehicle speed significantly affected acceptance rates and safety margins in both simulators. Conclusion: For a common decision task, the results obtained in virtual environments deviate from those in a nonvirtual test bed. The consistency of differences indicates that restrictions apply when predicting real-world behavior based on VR studies. In particular, the higher susceptibility to speed effects warrants further investigation, since it implies that differences in perceptual processing alter experimental outcomes. Objective: To contribute to the validation of virtual reality (VR) as a tool for analyzing pedestrian behavior, we compared two types of high-fidelity pedestrian simulators to a test track. Background: While VR has become a popular tool in pedestrian research, it is uncertain to what extent simulator studies evoke the same behavior as non-virtual environments. Method: An identical experimental procedure was replicated in a CAVE automatic virtual environment (CAVE), a head-mounted display (HMD), and on a test track. In each group, 30 participants were instructed to step forward whenever they felt the gap between two approaching vehicles was adequate for crossing. Results: Our analyses revealed distinct effects for the three environments. Overall acceptance was highest on the test track. In both simulators, crossings were initiated later, but a relationship between gap size and crossing initiation was apparent only in the CAVE. In contrast to the test track, vehicle speed significantly affected acceptance rates and safety margins in both simulators. Conclusion: For a common decision task, the results obtained in virtual environments deviate from those in a non-virtual test bed. The consistency of differences indicates that restrictions apply when predicting real-world behaviour based on VR studies. In particular, the higher susceptibility to speed effects warrants further investigation, since it implies that differences in perceptual processing alter experimental outcomes. Application: Our observations should inform the conclusions drawn from future research in pedestrian simulators, for example by accounting for a higher sensitivity to speed variations and a greater uncertainty associated with crossing decisions.