PaperlayerView exact revision on arXiv

Source-linked AI summary

Navigation Explainability for Object Avoidance with LLMs for Assistive Robots

Courtney Young, Yi Lian, Taery Kim, Sehoon Ha

arXiv:2610.28920v1cs.RO

TL;DR

BVI users may be unable to inspect their surroundings or infer why a robot guide dog changes course, and which explanations best support understanding and trust remains unclear. The paper presents a framework that combines obstacle information and planned future motion to generate natural-language explanations. In a user study, satisfaction strongly correlated with trust, and obstacle-plus-trajectory context at moderate frequency generally received the highest ratings for both.

  • Problem

    It remains unclear which natural-language explanations help BVI users understand robot behavior and build trust when they may not directly inspect the environment.

    Source: §I · p. 1

  • Method

    The framework supplies an LLM with detected obstacle information and future planner waypoints to generate real-time explanations of a robot guide dog’s navigation intent.

    Source: §I · p. 1

  • Results

    Explanation satisfaction was strongly positively correlated with trust, and obstacle-plus-trajectory context at moderate frequency generally received the highest ratings for both.

    Source: §I · p. 1

  • Takeaways & Limitations

    Users particularly valued explanations that were informative, complete, and useful to their goals.

    Source: §VII · p. 8

    The setup lacked spatial awareness of the user, YOLO missed some walls and support beams, and the in-application map sometimes had localization errors.

    Source: C. Limitations · p. 7

Abstract

from arXiv · show

Blind and visually impaired individuals face uncertainty when navigating unfamiliar areas. Whether a cane or a guide dog is used for navigation, this may prove to be challenging. To mitigate this challenge, we present the Context-Aware Navigation Explanation framework that explains guide dog movements for obstacle avoidance and navigation using a large language model. This system integrates perceptual context with planned future trajectories to generate natural-language explanations of robot movement. We deploy the system on a Unitree Go2 platform in indoor navigation scenarios and evaluate how different explanation context configurations affect user trust and explanation satisfaction. Results show a strong positive relationship between explanation satisfaction and trust, and suggest that combining obstacle information with future trajectory leads to the most effective explanations. These findings highlight the potential of context-aware language explanations to improve transparency and trust in assistive robot navigation.

I. INTRODUCTION

The introduction motivates explanations that help BVI users understand robot behavior and trust its decisions, then presents a framework that combines obstacle perception with planned motion. A user study reports that obstacle and trajectory context with moderate explanation frequency received the highest ratings for satisfaction and trust.

I. INTRODUCTION

  • The open question is which natural-language explanations help BVI users understand robot behavior and build trust when they may not directly inspect the environment.Guide-dog training costs and short service lives limit access to approximately 2% of the BVI population.Source: §I · p. 1
  • The framework gives an LLM detected obstacle information and future planner waypoints over a short temporal window to explain the robot’s current intent.The system generates real-time natural-language explanations of a robot guide dog’s navigation behavior.Source: §I · p. 1, Figure 1 · p. 1
  • A user study varies the type and amount of context to examine how explanation design affects satisfaction and overall system trust.The study investigates whether different LLM-generated comment content and quality correspond to differences in trust.Source: §I · p. 1
  • Satisfaction and trust were strongly positively correlated, while obstacle-plus-trajectory context at moderate frequency generally received the highest ratings on both.These ratings exceeded those for conditions without future trajectory or with higher explanation frequency; users particularly valued informative, complete explanations aligned with their goals.Source: §I · p. 1

II. RELATED WORK

Prior work frames assistive guiding as a combination of navigation, communication, and safety, with robot-guide systems extending physical guidance and interaction. Research also identifies trust, tailored communication, and real-time motion explanation as ongoing concerns.

II. RELATED WORK

  • Assistive guiding is presented as safety-critical, requiring reliable navigation and communication that help users understand and anticipate robot motion.The review covers BVI guide robots, trust barriers in assistive HRI, and foundation models for explainable navigation.Source: §II · p. 2

A. Robot Guide Dogs and Assistive Guiding for BVI Users

  • Robot-guide paradigms use mobile and quadruped platforms for physical guidance and richer interaction, while user studies favor semantic, personalized feedback and caution against excessive audio.This work frames guiding as spanning navigation, physical interaction, and safety [7].Source: A. Robot Guide Dogs and Assistive… · p. 2

B. Trust and Communication in Assistive HRI

  • BVI studies link trust to consistent behavior and clear communication, while real-time explanations of motion decisions under planning constraints and dynamic scenes remain an open challenge.Legible motion can convey intent but may remain ambiguous when environmental constraints cause unexpected deviations,.Source: B. Trust and Communication in… · p. 2

C. Explainable Navigation

The paper places real-time navigation explanations within explainable navigation research and proposes a framework that combines obstacle perception with planned trajectory information. Its navigation context represents recent environmental observations alongside the robot’s intended motion.

C. Explainable Navigation

  • Explainable navigation seeks to help users understand robot actions and anticipate future behavior, including when behavior differs from expectations.Source: C. Explainable Navigation · p. 2

III. CONTEXT-AWARE NAVIGATION EXPLANATION FRAMEWORK

  • The framework gives an LLM obstacle information and future navigation intent to generate real-time explanations of the robot’s motion.Source: §III · p. 2, Figure 2 · p. 3

A. Navigation Context Representation

B. LLM-Based Explanation Generation

The explanation module converts navigation context into text for an LLM, which generates concise natural-language descriptions. Prompts specify how to use obstacle and route information when describing navigation behavior.

B. LLM-Based Explanation Generation

C. Explanation Timing and Context Aggregation

Explanations are generated periodically from aggregated perception and trajectory context rather than at every sensor update. The framework runs alongside the robot’s navigation system during guidance.

C. Explanation Timing and Context Aggregation

IV. USER STUDY

The user study evaluates how navigation-context configurations affect perceived explanation quality and trust. It uses a within-subject design with conditions that vary the information and feedback frequency supplied to the LLM.

IV. USER STUDY

  • The study evaluates LLM-generated explanations during robot-assisted navigation and measures perceived explanation quality and trust in the robotic guide dog system.Source: §IV · p. 4

A. Study Design

  • The study uses a within-subject design to examine whether richer navigation context is associated with clearer, more informative, and more trustworthy explanations.Source: A. Study Design · p. 4
  • Participants experience multiple conditions that vary parameters used to construct the navigation context supplied to the LLM.Source: A. Study Design · p. 4
  • C1 provides obstacle information, while C2 and C3 add navigation information; C3 also uses a longer frame history and higher feedback frequency.Source: A. Study Design · p. 4

B. Participants

The study recruited 12 participants, including two BVI participants, and tested a quadrupedal guide robot on indoor routes with obstacles and turns. Participants completed six trials across three conditions and two paths, then rated trust and explanation quality.

B. Participants

  • The 12 participants included two BVI participants; the other 10 were sighted and wore eye masks during the experiments.The study protocol received IRB approval, and all participants provided informed consent.Source: B. Participants · p. 4

C. Experimental Setup

  • The study used a Unitree Go2 with a harness, spoken explanations, an RGB-D camera with YOLOv8n detection, and Nav2-planned waypoints.Testing took place in an indoor hallway with chairs, pedestrians, and turns along two predefined routes traversed in both directions.Source: C. Experimental Setup · p. 4, Figure 3 · p. 5

D. Procedure

  • After familiarization without explanations, participants followed the robot on predefined paths and completed six counterbalanced trials across three conditions and two paths.They completed a questionnaire after each trial.Source: D. Procedure · pp. 4–5

E. Measures

  • Post-run questionnaires used five-point Likert ratings for trust and explanation quality, with open-ended comments available to capture additional preferences or concerns.Trust items covered reliability, predictability, dependability, responsiveness, and communication; explanation items covered understanding, satisfaction, detail, completeness, usefulness, and informativeness.Source: E. Measures · p. 5

V. RESULTS

Trust and explanation satisfaction were strongly positively associated, while explanation condition significantly affected both measures. C2 received stronger satisfaction ratings than C3, with trust also trending higher for C2, though that trust difference did not survive correction.

A. Participant-Level Analysis

  • r = 0.949, p < 0.05 for the positive association between TPS-HRI trust and ESS satisfaction across participants; the association was similarly strong across conditions.Fig. 4 presents aggregated and per-condition correlations.Source: A. Participant-Level Analysis · p. 5, Figure 4 · p. 5
  • F(2, 22) = 3.951, p = 0.0342 for trust and F(2, 22) = 5.629, p = 0.0106 for satisfaction, showing significant condition effects on both measures.The reported generalized effect sizes were η2 g = 0.0409 for TPS-HRI and η2 g = 0.066 for ESS; the authors note the modest sample size (n=12). Fig. 5 shows mean scores.Source: A. Participant-Level Analysis · p. 6, Figure 5 · p. 6
  • C2 satisfaction (M = 3.74) exceeded C3 (M = 3.26; padj = 0.038), while C2 trust (M = 3.81) exceeded C3 (M = 3.49) without surviving Bonferroni correction (padj = 0.063).C2 also exceeded C1 in satisfaction, but that difference was not statistically significant after correction.Source: A. Participant-Level Analysis · p. 6

B. Item-Level Analysis

Item-level analyses found condition differences across trust and explanation items, with C2 generally receiving higher ratings. The reported comparisons particularly favor C2 on usefulness, satisfaction, completeness, detail, dependability, and appropriate information.

B. Item-Level Analysis

  • The item-level Friedman analysis found significant differences across conditions for trust- and explanation-related items (p < 0.05).Fig. 6 plots mean trust and explanation-satisfaction item scores across C1–C3 and identifies significant items.Source: B. Item-Level Analysis · p. 6, Figure 6 · p. 6
  • Satisfaction was rated significantly higher in C2 than C3 (p = 0.016, padj = 0.046); usefulness was also among items with significant p values below 0.05 and rated higher in C2 than other conditions.Source: B. Item-Level Analysis · p. 6, Figure 6 · p. 6
  • Completeness and sufficient detail favored C2 over C1, while dependability and appropriate information also had higher C2 ratings than C1.The reported adjusted p-values were 0.046 for completeness, 0.058 for sufficient detail, 0.134 for dependability, and 0.091 for appropriate information.Source: B. Item-Level Analysis · pp. 6–7

VI. DISCUSSION

Participants valued actionable obstacle guidance and clear, informative explanations, while BVI participants and sighted participants expressed differing preferences around distance descriptions and update frequency. BVI participants also described the robot as a trusted team partner when its comments sounded definitive.

A. General Discussion

  • Participants preferred obstacle detection paired with an actionable directive, which helped them prepare for the robot’s next movement.Participants reported increased trust after the robot carried out the described action.Source: A. General Discussion · p. 7

B. BVI Participant Preferences and Suggestions

  • BVI participants and most participants favored general distance descriptions, while some sighted participants preferred exact distances; some suggested expressing distance in steps.BVI participants did not report the cognitive overload that sighted participants observed with C3’s high-frequency comments, and they valued obstacle identification and hearing about obstacles at different distances.Source: B. BVI Participant Preferences and… · p. 7
  • BVI participants described the robot as a team partner and said its definitive, non-vague comments made it feel more trustworthy.One participant suggested that the robot introduce itself to make the team relationship more personal.Source: B. BVI Participant Preferences and… · p. 7

C. Limitations

The conclusion reports that explanation satisfaction is strongly correlated with trust and that obstacle-plus-trajectory explanations at moderate frequency tended to receive higher ratings. It also identifies system and explanation limitations and outlines future work.

C. Limitations

  • The robot lacked spatial awareness of users, missed walls and support beams that some participants collided with, and sometimes localized poorly in uniform spaces.Explanation errors in conditions C2 and C3 also incorrectly indicated a turn ahead and lowered one participant’s satisfaction.Source: C. Limitations · p. 7

VII. CONCLUSION

  • The framework uses LLMs to generate real-time explanations for a robotic guide dog by combining obstacle information with future navigation trajectory.Source: §VII · p. 7
  • Satisfaction was strongly correlated with trust, while obstacle-plus-trajectory explanations at moderate frequency tended to rate higher than no-future-trajectory or higher-frequency conditions.The study also found that users valued explanations that were informative, complete, and useful to their goals.Source: §VII · pp. 7–8
  • Users valued explanations that were informative, complete, and useful to their goals.Source: §VII · p. 8
  • The authors highlight context-rich, actionable explanations as important when assistive-navigation users rely on the robot to understand the environment.Source: §VII · p. 8
  • Future work includes haptic feedback, emergency alerts, safety-aware control, and larger BVI studies of long-term trust and real-world usability.Source: §VII · p. 8
Loading 2610.28920v1…