Evaluating the Viability of Additive Models to Predict Task Completion Time for 3D Interactions in Augmented Reality

TL;DR

This work investigates the viability of using Keystroke-Level Model (KLM)-style additive models to predict task completion time ($TCT$) for 3D augmented reality interactions across multiple input modalities. It builds a framework where each interaction is decomposed into movement time ($MT$) and confirmation operators ($CO$), and then sums these atomic times to forecast $TCT$ for both a simple menu selection task and a complex manipulation task, using six input modalities. By validating several literature-based $MT$ and $CO$ models (e.g., distal pointing with $ID_{ANG}$, gaze-based timing, and hand interaction with depth change), the paper demonstrates that additive models can predict absolute and relative performance with reasonable accuracy (often within 20%), though certain modalities (notably ControllerBlink) exhibit larger errors likely due to fatigue or tracking issues. The findings suggest additive 3D predictive models are a promising tool for AR design, enabling rapid comparison of input modalities and informing design decisions, while highlighting the need to tailor operator times to device and task specifics for robust deployment.

Abstract

Additive models of interaction performance, such as the Keystroke-Level Model (KLM), are tools that allow designers to compare and optimize the performance of user interfaces by summing the predicted times for the atomic components of a specific interaction to predict the total time it would take to complete that interaction. There has been extensive work in creating such additive models for 2D interfaces, but this approach has rarely been explored for 3D user interfaces. We propose a KLM-style additive model, based on existing atomic task models in the literature, to predict task completion time for 3D interaction tasks. We performed two studies to evaluate the feasibility of this approach across multiple input modalities, with one study using a simple menu selection task and the other a more complex manipulation task. We found that several of the models from the literature predicted actual task performance with less than 20% error in both the menu selection and manipulation study. Overall, we found that additive models can predict both absolute and relative performance of input modalities with reasonable accuracy.

Figures (3)

• Figure 1: Total predicted and actual time with confirmation operators in seconds for all participants for all input modalities for the menu selection study.
• Figure 2: Total predicted and actual time for all participants for all input modalities in the manipulation study.
• Figure 3: Total predicted vs. actual times without confirmation operators for each of the three phases of a manipulation trial.