PopNavShift: Stress-Testing Social Navigation under Behavioral Population Shift

Summary

Abstract Summary

Social-navigation algorithms are often evaluated under a fixed pedestrian-behavior distribution, despite substantial variation in pedestrian responses to robots across individuals and social contexts. We introduce PopNavShift, a matched simulation framework for stress-testing social-navigation strategies under pedestrian population shifts. PopNavShift constructs population-conditioned pedestrian motion profiles by prompting Gemini 3.7 Flash with 600 synthetic persona records from MatrAIx Persona 1M and deterministically mapping the responses into bounded motion parameters. It then compares three representative navigation strategies, reactive avoidance, early yielding, and reciprocal collision avoidance, across eight population conditions and 7,488 matched robot runs. In a matched intervention on the same 202 personas, changing only time pressure reverses 8.6% of controller rankings based on robot travel time, but 22.4% based on mean pedestrian delay and 23.9% based on worst-decile delay. Across population conditions, this sensitivity is greater for pedestrian burden than for robot travel time and increases in spatially constrained settings; the same qualitative pattern persists under a second pedestrian dynamics model. These findings support evaluating navigation strategies across behavioral populations using both robot performance and pedestrian burden.

Key Points

Social-navigation algorithms are often evaluated under a fixed pedestrian-behavior distribution, despite substantial variation in pedestrian responses to robots across individuals and social contexts.
We introduce PopNavShift, a matched simulation framework for stress-testing social-navigation strategies under pedestrian population shifts.
PopNavShift constructs population-conditioned pedestrian motion profiles by prompting Gemini 3.7 Flash with 600 synthetic persona records from MatrAIx Persona 1M and deterministically mapping the responses into bounded motion parameters.
It then compares three representative navigation strategies, reactive avoidance, early yielding, and reciprocal collision avoidance, across eight population conditions and 7,488 matched robot runs.
In a matched intervention on the same 202 personas, changing only time pressure reverses 8.6% of controller rankings based on robot travel time, but 22.4% based on mean pedestrian delay and 23.9% based on worst-decile delay.

Method

To operationalize the matched population-replay design, PopNavShift separates the construction of behavioral populations from pedestrian dynamics and robot control (Fig. 1 ). Eight fixed robot-encounter probes, each describing a standardized navigation situation, elicit structured responses from each persona, which a deterministic mapping converts into seven bounded motion controls.

Experimental Design

We evaluate PopNavShift on five types of physical encounters: a bidirectional sidewalk, a head-on encounter, a crossing, a bottleneck, and a static obstruction. The scenario grid varies available width (1.8, 2.4, 3.0, and 4.2 m), pedestrian density (0.06, 0.12, and 0.20 persons/m 2 ), and robot speed (0.60 and 1.00 m/s).

Head-on encounters contain one pedestrian and use the lowest density, yielding 104 valid physical cells rather than a full Cartesian product.

Experimental Design — PopNavShift: Stress-Testing Social Navigation under Behavioral Population Shift — Fig. 3: Changes in pedestrian delay under behavioral population shift. Entries show shifted minus reference delay over the same 104 physical cells. The outlined row shows the matched time-pressure intervention on the same personas. White separators distinguish the matched intervention, source-field cohorts, the mixed population, and probe-defined cohorts.

Results

The main experiment contains 7,488 robot runs and 2,496 matched runs without a robot, with paired pedestrian delay available for nearly all records. Changing time pressure from No rush / Moderate to Deadline alters at least one categorical response for 201 of the 202 matched personas. This shift changes controller comparisons more often for pedestrian burden than for robot travel time.

Mean-delay orderings reverse in 22.4% of comparable cells and worst-decile-delay orderings in 23.9%, compared with 8.6% for robot travel time.

Sources and Demonstrations

Figures are reproduced from Tan et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.