Bringing Reality into Virtual Reality: Safety Behavior Assessments in Automatically Generated Virtual Construction Environments

Current approaches to evaluating measures and interventions in construction safety struggle to capture the impact on worker behavior. Field studies are hazardous or lagging, while virtual testbeds often measure intentions and not behavior. Virtual environments also rely on manually and artificially made-up scenarios, which result in a high workload and limit real-world relevance. Moreover, existing works are primarily static environments without collaboration between workers or equipment.

This thesis investigates how virtual environments can utilize existing knowledge to (a) reduce workload in creating virtual scenes, (b) measure safety behavior instead of intention, (c) incorporate social dynamics through role-specific collaboration, and (d) ultimately result in relevant scenarios where outcomes in the virtual environments can demonstrate efficacy of the intervention or measure under investigation.

A framework was developed in six design science research (DSR) cycles. It combines (a) an ontology-based representation of safety knowledge, (b) Artificial Intelligence (AI) methods to align risk information, (c) a data model to automate generating the environments directly from this knowledge, (d) multi-user and equipment integration for collaboration, and (e) an energy-based metric weighting close calls to measure safety behavior based on hazardous interactions rather than intentions. Each cycle included case studies investigating the impact of different interventions, including safety training, safe path guidance technology, or a communication technology.
Findings show that existing knowledge from ontologies can be transformed into
interoperable environments, enabling automatic scenario generation from real-world
hazards while ensuring content validity and relevance. The energy-based indicator
allows for assessing behavioral safety performance, while multi-user experiments
enable collaborative and cooperative aspects. These contributions provide safe
methods for behavioral assessments in virtual environments to demonstrate the
efficacy of safety interventions, bringing reality into virtual worlds.

Future work should investigate (a) the use of generative AI to automate scenario
creation from unstructured knowledge and (b) subject-free simulations to predict
hazards, bottlenecks, and layout risks. Together, these could evolve the framework
from a feasibility framework into a decision-support tool for safety management.