Avatar fidelity and Co-presence in a collaborative VR educational task: a secondary analysis of the HuTwin dataset

In plain English

A re-analysis of an open dataset of 40 people working in pairs on a shared cooking task in VR, comparing a stylised chef avatar with a more human-like humanoid one at 0 ms and 500 ms delay. The more human-like avatar raised co-presence ratings by 1.12 points with no delay, but the advantage shrank to a non-significant 0.38 once 500 ms of latency was added, while avatar type made no difference to how fast people finished the task and the 500 ms delay slowed them by about 17.4%. The practitioner takeaway is blunt: realism buys felt presence, not productivity, and latency eats the realism benefit before it eats the task.