Goal-Expressive Movement for Social Navigation: Where and When to Behave Legibly

Abstract

Robots often need to communicate their navigation goals to humans to assist observers in anticipating the robot's future actions. Enabling observers to infer where the robot is going from its movements is particularly important as robots begin to share workplaces, sidewalks, and social spaces with humans. We can use legible motion, or movements that use intentional inefficiencies to indicate the robot's goal, to communicate this information non-verbally, which is especially valuable in loud environments or ones where humans are multitasking.

However, while legibility has enabled robots to communicate their intentions when performing manipulation tasks, we have found that applying legibility to social navigation highlights differences between these contexts. In social navigation, robots travel over greater distances, and it is more likely that observers cannot see all parts of a robot's path. The heading of a robot and its perceived "gaze" matter strongly in navigation, and it is important to account for this signal when creating legible paths. And in navigation, there are likely to be many goals under consideration at varying depths from the start location, and the robot must often pass through ambiguous areas where it is difficult to telegraph a final destination. An ideal legible motion algorithm for social navigation would enable robots to be expressive when that effort is relevant to the target goal, but not be overly expressive when that would be unhelpful or misleadingly indicate other goals.

To explore and improve these aspects of legibility in social navigation, we first conducted a 300-participant virtual user study exploring observer-aware legibility, a formulation of legibility we created to ensure that legible movements are made when within vision for a target observer. In this study, we also developed several metrics that assess when and how well legible paths successfully communicate their goal. We found that observer-aware legibility was particularly effective at ensuring observers with poor views were better able to see and understand the robot's path before it arrived. We also discovered that an observer's view of the environment relative to the robot and goals has a major impact on their interpretations of the robot's goals.

By analyzing the detailed data we collected in this experiment, we found that observer reactions to legible motion differed from expectations. We also found that mathematically, modeling what information observers took from a legible path was fundamentally different than the mathematical model used for generating legible paths. This required a new model, which we call understanding. We highlight three ways in which modeling observer understanding differs from the process of generating legible paths. One of our most notable observations was that even in the absence of ambiguity, later moments in the path are found to have a larger impact on understanding than earlier moments, a finding which is contrary to the traditional goal of being legible as early as possible.

Using these findings, we designed an improved algorithm for legibility for social navigation, adding crucial adjustments to enable the principles of legibility to scale to scenarios with increasingly many goals at varying depths and arrangements, which was previously unsupported. This was achieved by using the insight that observers only take a subset of goals under consideration at each point, and that in ambiguous areas where other goals are more likely than the true destination, it is most appropriate to focus efforts on avoiding indicating the robot's goal incorrectly.

We designed a user study to investigate: is it truly more effective to be legible as early as possible, or later in the path when it may be more informative? To test this, we conducted a 32-person user study with a Fetch robot moving between six goals, both with and without the presence of a distractor task. We found that regardless of distraction, if there was an ambiguous zone early the robot had to pass through that made communicating intent difficult, paths that were legible later in the path gave human observers a larger period of time where they were confident in the robot's goal than those that attempted to be legible early in the path. We also discuss how our findings suggest that legibility is physically constrained to certain areas of a given scenario, and it is most effective to telegraph intent within those zones, overall clarifying where and when it is most effective to employ legible motion in social navigation.