A Networked Desktop Virtual Reality Setup for Decision Science and Navigation Experiments with Multiple Participants

Research on spatial cognition and navigation typically studies the spatial decision-making (e.g., turning left or right at an intersection) and mental representation of individuals in real and virtual environments^1,2. The advantages of virtual reality (VR) include the prevention of ethical and safety issues (e.g., during a dangerous evacuation³), the automatic measurement and analysis of spatial data⁴, and a balanced combination of internal and external validity^5,6,7. For example, Weisberg and colleagues extended previous research on individual differences in spatial knowledge acquisition by demonstrating that spatial tasks in VR can provide an objective behavioral measure of spatial ability⁸. This study also suggested that the navigation behavior in VR approximates real-world navigation because the virtual environment was modeled after the university campus used by Schinazi and colleagues⁹ (see also the study of Ruddle and colleagues¹⁰). VR has also been applied to psychotherapy¹¹, clinical assessment¹², consumer behavior¹³, and surgery^14,15. However, most VR systems lack proprioceptive and audio feedback that may improve presence and immersion^16,17,18,19, require training with the control interface^20,21,22, and lack social cues. Indeed, people in the real world often move in groups²³, avoid or follow other people^3,24, and make decisions based on social context^25,26.

At the same time, research on crowd behavior often focuses on emergent characteristics of crowds (e.g., lane formation, congestion at bottlenecks) that are simulated on a computer or observed in the real world. For example, Helbing and colleagues used a combination of real-world observations and computer simulations in order to suggest improvements to traffic flow in an intersection by separating inflow and outflow with physical barriers and placing an obstacle in the center²⁷. Moussaïd and colleagues used a heuristics-based model to study high-density situations during a crowd disaster²⁸. This approach suggested improvements to an environmental setting for mass events in order to avoid crowd disasters. With the aid of an existing open source framework, the implementation of such simulations could be relatively easy. SteerSuite is an open source framework that allows users to simulate steering algorithms and crowd behavior easily by providing tools for facilitating, benchmarking, and testing²⁹. This framework can provide the core of an agent's navigation rationale, which is critical for successful crowd simulation. In addition, Singh and colleagues demonstrated a single platform that combines a variety of steering techniques³⁰. While researchers can propose design interventions using such simulations, they are rarely validated with human participants in a controlled setting. Controlled experiments are rare in crowd research because they can be difficult to organize and dangerous to the participants.

VR has been employed to investigate social behavior using simple and complex virtual environments with one or more computer-simulated agents. In the study of Bode and colleagues^31,32, the participants were asked to evacuate a simple virtual environment from a top-down perspective among several agents and found that exit choice was affected by static signage and motivation. Presenting participants with a more complex environment from a first-person perspective, Kinateder and colleagues found that the participants were more likely to follow a single computer-simulated agent during the escape from a virtual tunnel fire²⁵. In a complex virtual environment with multiple agents, Drury and colleagues found that the participants tended to assist a fallen agent during an evacuation when they identified with the crowd²⁶. Collectively, these findings suggest that VR can be an effective way of eliciting social behaviors, even with computer-simulated agents. However, some crowd behaviors may only be observed when there is a realistic social signal (i.e., when the participants are aware that the other avatars are controlled by people³). In order to address this shortcoming, the present protocol describes a method for conducting controlled experiments with multiple users in a networked VR setup. This approach has been employed in a recent study by Moussaid and colleagues in order to investigate the evacuation behavior of 36 networked participants³.

Research on networked VR has focused on topics unrelated to navigation strategies^33,34 and/or relied on existing online gaming platforms such as Second Life. For example, Molka-Danielsen and Chabada investigated evacuation behavior in terms of exit choice and spatial knowledge of the building using participants recruited among existing users of Second Life³⁵. While the authors provide some descriptive results (e.g., visualizations of trajectories), this study had difficulties with participant recruitment, experimental control, and generalization beyond this specific case. More recently, Normoyle and colleagues found that existing users of Second Life and participants in a laboratory were comparable in terms of evacuation performance and exit choice and different in terms of self-reported presence and frustration with the control interface³⁶. The findings from these two studies highlight some of the challenges and opportunities afforded by online and laboratory experiments. Online studies are capable of drawing from a much larger and motivated population of potential participants. However, laboratory studies allow for more experimental control of the physical environment and potential distractions. In addition, online studies may pose some ethical concerns regarding data anonymity and confidentiality.

As a networked desktop VR laboratory, the Decision Science Laboratory (DeSciL) at ETH Zürich is primarily used to study economic decision-making and strategic interactions in a controlled environment. The technical infrastructure at the DeSciL consists of hardware, software for laboratory automation, and software that supports the multi-user desktop VR setup. The hardware includes high-performance desktop computers with Microsoft Windows 10 Enterprise operating system, control interfaces (e.g., mouse and keyboard, joysticks), headphones, and eye trackers (Table of Materials). All client computers are connected with Ethernet of one gigabit per second to the university network and the same network file share. There is no visible delay or lag when there are 36 clients connected. The number of frames per second is consistently above 100. The experiments are also managed and controlled with laboratory automation software based on Microsoft PowerShell (i.e., PowerShell Desired State Configuration and PowerShell Remoting). All relevant steps of the protocol are preprogrammed with PowerShell scripts called Cmdlets (e.g., Start-Computer, Stop-Computer). During the experiment, these scripts can be executed simultaneously and remotely on all client computers. This type of laboratory automation ensures an identical state of the client computers, reduces potential errors and complexity during scientific testing, and prevents researchers from having to perform repetitive manual tasks. For the navigation experiments, we use the Unity game engine (<https://unity3d.com/>) in order to support the development of 2D and 3D environments for multi-user, interactive desktop VR. The 36 client computers are connected to a server via an authoritative server architecture. At the start of every experiment, each client sends an instantiation request to the server, and the server responds by instantiating an avatar for that user on all of the connected machines. Each user's avatar has a camera with a 50 degrees field of view. Throughout the experiment, the clients send user' input to the server, and the server updates the movement of all of the clients.

In the physical laboratory, each computer is contained in a separate cubicle within three semi-independent rooms (Figure 1). The overall size of the laboratory is 170 m² (150 m² for experiment room and 20 m² for control room). Each of these rooms is equipped with audio and video recording devices. Experiments are controlled from a separate adjacent room (i.e., by providing instructions and initiating the experimental program). From this control room, the experimenters can also observe the participants in both physical and virtual environments. Together with the Department of Economics at the University of Zürich, the DeSciL also maintains the University Registration Center for Study Participants, which was implemented based on h-root³⁷.

Although similar systems have been described in the literature³⁸, the DeSciL is the first functional laboratory that is suitable for multi-user desktop VR experiments on navigation and crowd behavior to our knowledge. Here, we describe the protocol for conducting an experiment in the DeSciL, present representative results from one study on social navigation behavior and discuss the potential and limitations of this system.