PaperlayerView exact revision on arXiv

Source-linked AI summary

Rethinking Legibility in Social Robot Hallway Navigation: Impact of Intent Representation and Human Distraction

Pranav Goyal, Andrew Stratton, Christoforos Mavrogiannis

arXiv:2609.40158v1cs.ROcs.HC

TL;DR

In constrained social navigation, pedestrians need to coordinate around near-term conflicts, while much legible-motion work represents intent as a destination. The paper compares intent representations in two controlled hallway studies and finds that passing-side strategies can reduce human acceleration, including under distraction, though subjective ratings under distraction did not distinguish the algorithms.

  • Problem

    Destination-based intent is common in legible motion, but constrained social navigation requires pedestrians to coordinate conflict resolution rather than infer final goals.

    Source: §1 · p. 1, §3.3 · p. 7

  • Method

    The authors compare alternative intent representations within a shared model predictive control framework across two controlled hallway studies, including a study of distraction.

    Source: §1 · p. 3

  • Results

    Study 1 found passing-side signaling reduced Human AA versus the non-legible baseline; in Study 2, SM reduced Human AA versus NL, while subjective ratings did not significantly distinguish algorithms under distraction.

    Source: §7 · pp. 20–21

  • Takeaways & Limitations

    The findings support designing social navigation around clear near-term motion and adaptation to pedestrians, and evaluating both perceptions and observed coordination.

    Source: §7 · p. 21

    The studies used scripted, one-on-one encounters in a controlled laboratory hallway, and legibility remained intertwined with passing-side norms.

    Source: §6.6 · p. 20

Abstract

from arXiv · show

We focus on legible robot motion generation in social navigation settings. Legibility in human-robot interaction (HRI) is often described as the property of robot motion that enables an observer to confidently infer the robot's intent. While mature frameworks exist for generating legible motion in front of static observers, social robot navigation presents a new challenge: the robot must clearly convey its intent while ensuring human safety in dynamic pedestrian environments where human attention is often divided. With the goal of enabling robots to generate legible motion in dynamic and constrained spaces, we investigate how the choice of representation and the level of human attention shape navigation performance and human impressions. Focusing on the ubiquitous and demanding scenario of hallway navigation, we conduct two controlled user studies involving alternative legibility formulations implemented within a shared model predictive control framework. Study 1 (N = 45) investigates the role of intent representation, showing that passing-side legibility, particularly when adaptively updated, leads to smoother human motion and is perceived as more competent and less mentally and physically demanding than destination-based and non-legible baselines. Study 2 (N = 45) examines the effect of pedestrian attention, demonstrating that legible motion allows for smooth human motion even under distraction, even if this is not consistently reflected in subjective ratings. Together, these findings suggest that effective legible motion in social robot navigation benefits from interaction-level intent representations that support coordination, with some effects persisting even when human attention is divided. Code is available at https://github.com/fluentrobotics/Legible_MPPI.

1 Introduction

Legibility can help people interpret robot intent, but hallway navigation requires coordination among moving actors, not just inference by passive observers. The paper compares intent representations and tests whether distraction changes legible motion’s effects in two hallway studies.

1 Introduction

  • Earlier legibility work framed motion as a way to infer robot intent, with later studies examining viewpoint and transfer across domains,.Legible motion was also linked to collaboration, safety, and efficiency in shared tasks,.Source: §1 · pp. 1–2
  • In constrained navigation, pedestrians need to resolve conflicts rather than infer a robot’s final destination, while prior work often assumes passive, attentive observers.The paper identifies embodied interaction, hallways, and divided attention as underexplored settings.Source: §1 · p. 2
  • The paper asks how intent representation affects legibility in constrained navigation and how human attention shapes its effects during interaction.The authors argue destination-based intent is challenging to define in dynamic, crowded, constrained settings.Source: §1 · p. 2
  • Two comparative hallway studies embed alternative legibility formulations in a shared model predictive control framework, with 45 participants in each study.Study 1 examines intent representation; Study 2 treats distraction as an experimental variable and analyzes behavior and perception.Source: §1 · pp. 2–3
  • The studies compare destination-based and interaction-level intent, isolate representation and adaptability, and assess human navigation and subjective impressions.Measures include perceived competence, comfort, and effort, plus trajectory smoothness and human acceleration.Source: §1 · p. 3
  • Compared with an earlier four-page report, this manuscript adds formal foundations, objective trajectory-based measures, expanded discussion, and fuller statistical reporting.The added measures include conflict-resolution time and robot responsibility.Source: §1 · p. 3

2 Related Work

Related work moves from social navigation that models people as obstacles toward interaction-aware planning and robot motion that communicates intent. Prior legible-navigation studies vary in their intent representations, settings, and evaluation, leaving their comparative effects on people insufficiently understood.

Social robot navigation

  • Social navigation methods increasingly model mutual adaptation and seek efficient, comfortable motion, but often frame interaction as the robot inferring human intent.Earlier approaches treated people as moving obstacles; newer methods include learned policies and model-based planners [13], [67].Source: §2 · pp. 3–4

Legible motion

  • Legible motion encodes robot intent in task-oriented movement and is often evaluated through passive observation, which omits pedestrians’ reciprocal adaptation,.Embodied hallway encounters can involve misunderstandings and stand-offs.Source: §2 · p. 4

Legible motion in social navigation

  • Legible social-navigation methods represent intent through destinations or interaction cues such as passing side, but their comparative effects in embodied encounters remain insufficiently understood.Prior examples include online signal adaptation, Social Momentum, and rapid passing-side commitment [11].Source: §2 · p. 4

Hallway navigation

  • Hallway studies have examined avoidance strategies, passing distance, yielding, and gaze; this work adds intent expressiveness to that navigation setting.Earlier studies found that overly close passes invaded personal space, while excessively wide detours appeared inefficient.Source: §2 · p. 5

3 Foundations

This section formalizes legible social robot navigation and frames questions about intent representation and pedestrian distraction. It distinguishes inferring a destination from coordinating how pedestrians resolve encounters.

3.1 Domain

  • The robot’s navigation domain includes obstacles and co-navigating humans, with a destination goal subject to human safety and comfort.Source: §3.1 · p. 5
  • The framework models intent inference from partial robot trajectories and treats context as information that conditions inference, including maps, past behavior, or task cues.In the implementation, inference uses candidate intents and the observed trajectory’s cost.Source: §3.1 · p. 5
  • Legibility maximizes the posterior probability of the intended goal, whereas predictability maximizes the likelihood of motion given a goal; the two coincide when efficient motion toward that goal is incompatible with other intents.A corridor is given as an example where paths toward opposite endpoints are incompatible.Source: §3.1 · p. 6
  • MPC integrates legibility with safety, efficiency, and obstacle avoidance through a composite cost, with earlier legibility evidence weighted more heavily.This weighting encourages prompt intent disambiguation.Source: §3.1 · p. 6

3.2 Modeling the Pedestrian’s Belief

  • The pedestrian belief model represents inattention through λ: higher values produce more diffuse, uncertain beliefs, while values approaching zero produce confident inferences.Source: §3.2 · pp. 6–7

3.3 Challenges in Generating Legible Motion for SRN

  • The study asks whether destination-based intent scales to constrained navigation, where pedestrians may instead need to infer passing side and may change their intent during an encounter.A pedestrian may head toward an intermediate waypoint while resolving a conflict.Source: §3.3 · p. 7
  • The section asks how intent representation shapes legibility and how distraction affects navigation performance and human impressions, noting a potential trade-off with functional efficiency.Prior work assumes attentive observers; this work focuses on attention as a contextual factor.Source: §3.3 · p. 7

4 How does Intent Representation Shape Legible SRN

This study compares destination-based and passing-side intent representations for legible hallway navigation, including fixed and adaptive strategies. Its findings favor adaptive passing-side approaches for subjective ratings and human motion, while exploratory measures suggest earlier coordination under passing-side strategies.

4.1 Experiment Design

  • In a repeated-measures hallway study, 45 participants completed five counterbalanced algorithm trials while navigating head-on with the robot.Participants traversed the hallway to place stickers; a gong synchronized their turns with the robot’s movement.Source: §4.1 · p. 8, §4.2 · p. 9

4.2 Conditions

  • GL represents the robot’s destination, PL fixes a randomly assigned passing side, DPL adapts the side using predicted human motion, and SM adapts it through interaction-space angular momentum.NL is vanilla MPC without communicative intent; the strategies share a navigation framework but differ in how intent is represented and conveyed.Source: §4.2 · p. 9

4.3 Implementation

  • Controllers were tuned separately, and MPPI evaluated 250 trajectories over a 10-step horizon at 20 Hz.Human velocity was estimated from tracked positions at 0.05-second intervals for constant-velocity prediction.Source: §4.3 · p. 9

4.4 Measures

  • Human path irregularity and average acceleration measured motion, while RoSAS, NASA-TLX, and two Likert items captured impressions, workload, collision risk, and perceived legibility.The perceived-legibility item showed no significant differences across conditions; Human PI, L2, and NASA-TLX Performance, Effort, and Temporal Demand also had no significant effects.Source: §4.4 · p. 10, §4.6 · p. 12, §4.7 · p. 13

4.5 Hypotheses

  • The hypotheses predicted that legibility would improve impressions and human motion, passing-side intent would outperform destination intent, and dynamic adaptation would outperform fixed passing-side intent.The predicted outcomes included lower acceleration and workload and higher competence and comfort.Source: §4.5 · pp. 10–11

4.6 Analysis

  • Adaptive passing-side strategies (SM, DPL) were rated most competent, and DPL elicited the lowest discomfort; they also yielded smoother human motion than destination-based GL and non-legible NL.Across conditions, GL elicited higher workload and less fluent trajectories; overall results favored SM and DPL over NL and, in most cases, PL.Source: Table 2 · p. 12, §4.6 · pp. 12–13

4.7 Exploratory Analysis

  • Passing-side strategies (SM, DPL, PL) had the longest conflict-resolution times, indicating earlier human commitment to a passing side; SM and DPL also had the largest robot avoidance share.Conflict-resolution time was shortest under GL and NL, where comfortable clearance emerged only just before the closest encounter.Source: §4.7 · p. 14

5 How Human Distraction Impacts Benefits of Legible SRN

Study 2 tested whether legible navigation benefits persist when pedestrians are distracted. Objective coordination and smoothness measures continued to favor legible strategies, although subjective ratings did not reliably distinguish algorithms under distraction.

5.1 Study Design

  • Study 2 reused Study 1’s no-distraction data and added 45 participants to a distraction condition involving hallway navigation with a concurrent listening-and-phone comprehension task.Participants compared DPL, SM, and NL in a within-subjects design; the distraction task was piloted as manageable.Source: §5 · p. 14, §5.1 · pp. 14–16, §5.2 · p. 16

5.2 Analysis

  • Under distraction, Human AA was lower for DPL and SM than NL, with the NL–SM difference significant (p< .05); distraction did not significantly alter algorithm performance or impressions, so H4 was unsupported.Subjective measures did not differ significantly between algorithms under distraction.Source: §5.2 · p. 16

5.3 Exploratory Analysis

  • Competence and discomfort ratings had lower residual variance under distraction (1.24 vs. 1.01 and 0.71 vs. 0.46), with more uniform evaluations making algorithm differences harder to detect.The model comparisons favored a homoskedastic model (ΔAIC = 1.5; χ^2(2) = 2.47; p= .29).Source: §5.3 · pp. 16–17
  • Under distraction, users walked with significantly lower acceleration beside SM than NL (p< .05), although this smoother motion was not reflected in subjective ratings.Human AA residual variance increased from 0.041 without distraction to 0.072 under distraction.Source: §5.3 · p. 17
  • Conflict-resolution time was 2.35 s under SM, 1.78 s under DPL, and 0.81 s under NL; robot responsibility was 0.24, 0.23, and 0.20, respectively.All pairwise conflict-resolution-time differences and the legible-versus-NL robot-responsibility differences were significant (p< .01); SM and DPL did not differ in responsibility.Source: §5.3 · p. 17
  • Under distraction, SM outperformed NL in Human AA by Δ = 0.12, 0.19, and 0.25 at the pooled 25th, 50th, and 75th walking-speed percentiles; without distraction, a significant Δ = 0.17 appeared only at faster speeds.The three-way interaction was not significant; the planned contrasts showed a widening SM–NL gap with higher speed under distraction.Source: §5.3 · p. 17, Figure 6 · p. 18
  • The findings support context-sensitive tuning of the balance between legibility and efficiency, including future work on adapting λ to surrounding conditions.The section summary notes that objective benefits were more pronounced at higher walking speeds, despite no significant overall interaction for speed.Source: §5.3 · pp. 17–18

6 Discussion

The discussion argues that legibility in hallway navigation depends on how intent is represented and how pedestrians attend to the encounter. It favors adaptive, interaction-level signals while noting that social norms and the controlled study setting constrain interpretation.

6.1 Intent representation shapes legible motion

  • In the tested hallway setting, destination-based legibility was associated with higher workload and less smooth human motion, whereas adaptive passing-side strategies supported smoother coordination and higher competence ratings.Objective measures also indicated earlier pedestrian commitment to a passing side and a larger robot share of avoidance under passing-side strategies.Source: §6.1 · pp. 18–19

6.2 Adaptation strengthens legible motion

  • Adaptive strategies adjusted expressed passing intent to human motion, while static passing-side legibility had weaker and less consistent subjective benefits and could be perceived as rigid or unresponsive.The discussion frames pedestrians as interactive agents whose movements and emerging passing preferences can inform the robot’s behavior.Source: §6.2 · p. 19

6.3 Impact of social norms

  • Random-side PL performed better objectively than GL and NL despite deviating from the passing convention most, but adaptive strategies intertwine legible signaling with socially appropriate accommodation.The authors note that intent clarification, path-conflict resolution, and personal-space expectations are difficult to separate in this setting.Source: §6.3 · p. 19

6.4 Benefits persist under divided attention

  • Under distraction, objective coordination measures favored legible strategies even as subjective ratings became less sensitive; the authors suggest pedestrians may rely more on peripheral motion cues.Exploratory analyses found more pronounced benefits at typical and higher walking speeds, while Study 2 showed no significant subjective differences between algorithms.Source: §6.4 · pp. 19–20

6.5 Design implications

  • For constrained, hallway-like environments, the discussion recommends interaction-level intent representations, adaptive signaling, and treating human attention as an explicit contextual variable.The balance between functional objectives and interaction-level legibility is described as tunable within the shared control framework.Source: §6.5 · p. 20

6.6 Limitations and future work

  • The studies used scripted, one-on-one encounters in a controlled hallway and a fixed robot maximum speed of 0.33 m/s, so findings characterize a single operating point.The authors identify testing other encounter geometries, crowds, and speeds as future work.Source: §6.6 · p. 20

7 Conclusion

In head-on hallway encounters, passing-side signaling—especially when adaptive—supported smoother human motion and higher competence ratings than destination-based or non-legible approaches. Under distraction, SM still reduced Human AA relative to NL, although subjective ratings did not distinguish the algorithms.

7 Conclusion

  • In Study 1, passing-side signaling reduced Human AA relative to NL, while adaptive passing-side strategies received higher competence ratings; destination-based legibility elicited greater workload and Human AA than adaptive strategies.The conclusion states that near-term passing behavior was more useful for coordination than final destination alone in this setting.Source: §7 · pp. 20–21
  • In Study 2, SM reduced Human AA relative to NL under distraction, but subjective ratings did not significantly distinguish the algorithms.Source: §7 · p. 21
  • The studies support designing social navigation around clear near-term motion and adaptation to pedestrian movement, and evaluating both human perceptions and observed coordination.The conclusion calls for testing other encounter geometries and passing norms while directly measuring how pedestrians infer robot intent.Source: §7 · p. 21
Loading 2609.40158v1…