2026 ASEE Annual Conference & Exposition

Virtual reality (VR) training simulations are increasingly used in high-consequence settings such as semiconductor cleanroom operations. Although VR reduces the cost and risk of early proce-dural practice, scaling instruction remains difficult because proctors must continuously monitor learner progress, interpret where attention is directed, and intervene when errors or confusion oc-cur. This challenge becomes more acute when one instructor supervises multiple learners at the same time.
This paper presents a WebSocket-orchestrated framework for streaming multimodal data from multiple concurrent VR learners to a browser-based proctor dashboard without requiring the proctor to enter the headset experience. The framework streams egocentric video, gaze-derived fixation events, and step-level training-log events so that an instructor can view learner context, attention, and progress from a single interface.
We focus on the streaming architecture, runtime model, and scalability of the orchestration layer. Using synthetic clients and two RTX 3090-class machines, the system maintained sub-100 ms p95 round-trip latency through 32 concurrent learners under baseline gaze and training-log traffic (37 ms gaze and 70 ms training-log RTT at the highest baseline rate). Adding Meta Quest-like video streaming (480x270 JPEG at 8 fps) increased p95 latency only modestly (64–65 ms for gaze/training-log RTT and 35.7 ms for video ACK RTT). Continuous 10 Hz audio streaming rep-resents a worst-case stress condition and can dominate the server event loop at high concurrency, motivating batching, voice-activity gating, and selective worker offload for larger deployments.