Abstract
In the current work, we focus on systems that provide incremental directions and monitor the progress of mobile users following those directions. Such directions are based on dynamic quantities like the visibility of reference points and their distance from the user. An intelligent navigation assistant might take advantage of the user's mobility within the setting to achieve communicative goals, for example, by repositioning him to a point from which a description of the target is easier to produce. Calculating spatial variables over a corpus of human-human data developed for this study, we trained a classifier to detect contexts in which a target object can be felicitously described. Our algorithm matched the human subjects with 86% precision.
Introduction
Most automated navigation systems provide instructions based on simple triggers, such as proximity to a turn. However, human direction-givers are much more strategic, accounting for "felicity conditions"—the principle that a description is only useful if the listener is in a physical position to understand it. If a target object is obscured or positioned at an awkward angle, a human guide will often issue a repositioning command (e.g., "Turn right 90 degrees") before identifying the object. This research models the decision-making process of when the spatial relationship between a user and a goal is "good enough" to initiate a description.
Methodology and Spatial Modeling
The researchers utilized a virtual reality (VR) setup to collect human-to-human dialogue data. Using the OSU Quake 2004 corpus, which includes 15 recorded dialogues and 1,736 markable items, the authors built a Decision Tree Classifier to automate timing decisions.
As shown in Figure 4, several dynamic spatial features were extracted from VR logs to train the model:
- Angle (α): The absolute angle between the user’s view and the target.
- Distance (d): The physical distance from the user to the target.
- VisDistracts: The number of other visible objects in the scene.
- VisSemDistracts: The number of visible objects of the same category (e.g., other doors).
- Visible: A boolean indicating if the target is within the field of view (v).
Key Results and Insights
The decision tree correctly classified the timing of instructions with 86% precision, significantly outperforming a baseline model that triggered instructions based solely on visibility (70% accuracy).
Analysis of the model revealed several key insights into human navigation logic:
- Distractors as Bottlenecks: The most critical feature was
VisDistracts. If more than three objects are in view, humans typically delay descriptions to avoid confusion. - The "In Front" Bias: The model confirmed a strong preference for describing objects when the angle to the target is less than 33 degrees, aligning with human ease in perceiving the "in front" axis.
- Visibility vs. Felicity: Technical visibility does not equate to communicative effectiveness. The model identified specific "sweet spots" for instructions, such as an angle of ≤ 33° and a distance of ≤ 154 units in uncluttered scenes.
Conclusion and Future Work
This study demonstrates that effective navigational instruction requires strategic content planning based on spatial context rather than simple proximity triggers. While the current numerical thresholds are specific to the VR environment used, the underlying logic of monitoring for felicity conditions is broadly applicable. Future work will focus on integrating this timing logic with natural language generation systems and testing the model in more complex, high-density urban environments.
Summary updated