For You, Not for Everyone: Bystander-Aware Delivery of Critical Feedback in Robot Tutoring

Atikkhan Faridkhan Nilgar, Manuel Dietrich, Kristof Van LaerhovenarXiv 2026 · arXiv

TL;DR

Researchers developed a "bystander-aware" robot tutor that detects when a third party is nearby and adapts its delivery of critical feedback from spoken words to silent on-screen text. While the adaptive behavior significantly boosted the robot's perceived privacy awareness and social competence, it didn't conclusively reduce the learners' sense of embarrassment.

What Problem Does This Paper Address?

Social robots are moving out of the lab and into schools, homes, and hospitals—places where interactions rarely happen in total isolation. When a robot tutor gives a student critical feedback (e.g., "Your pronunciation was terrible"), doing so aloud in a shared space can be a major privacy violation and a social "faux pas."

Most HRI (Human-Robot Interaction) privacy research focuses on how data is collected or stored. This paper tackles a different angle: social communication privacy. How can a robot adapt its behavior in real-time to prevent sensitive information from leaking to bystanders?

Methods

The authors conducted a between-subjects study with 40 participants acting as Japanese language learners. The setup involved an Ohbot 2.2 robot head mounted on a custom 3D-printed torso with a smartphone display and a motorized turntable for body rotation.

The study compared two experimental conditions:

  1. Context-Invariant: The robot delivered critical feedback aloud, regardless of whether a bystander (a trained confederate) was present.
  2. Context-Adaptive: Upon reaching a feedback point, the robot performed a visual scan of the room. If it detected the bystander using its camera and a MediaPipe-based face/pose detection pipeline, it turned its body away from the bystander, gave the participant a spoken attention cue, and displayed the feedback silently on its screen.
Figure 3: Context-adaptive behavior from the participant’s viewpoint. From left to right, the robot tutor teaches pro- nunciation, scans for the bystander, then turns its body and displays feedback on-screen.

Crucially, the feedback was "scripted" and independent of actual performance (e.g., "I didn't know Japanese pronunciation could be butchered like that") to ensure the content remained constant across all participants.

Datasets

As this is a primary user study in HRI, the "dataset" consists of the qualitative and quantitative responses from the 40 adult participants (20 per condition).

  • Measures: Perceived privacy awareness (adapted from Tang et al.), Perceived Social Competence (SOC scale), and experienced embarrassment (adapted from Pitardi et al.).
  • Benchmarks: The study compares the novel adaptive behavior against a standard "invariant" baseline.

Key Insights and Ideas

  • Physicality as Privacy Signaling: The robot didn't need to explain why it was turning. Participants correctly inferred that the "looking around" behavior was a search for other people and that the subsequent turn was intended to keep the feedback private.
  • Privacy is a Social Skill: The fact that the adaptive robot was rated higher in "social competence" suggests that users view privacy-preserving behavior not just as a security feature, but as a sign of social intelligence.
  • The Embarrassment Paradox: Even though the feedback was delivered privately, embarrassment scores didn't drop as much as expected. Participant accounts suggest that simply performing a task poorly in front of a bystander is embarrassing, regardless of how the robot delivers the critique. The "social exposure" of the performance itself may outweigh the exposure of the feedback.
Comparison of the context-invariant (top) and context-adaptive (bottom) experimental setups.

Key Results

  • Privacy Awareness: The adaptive robot was rated significantly higher ($M = 5.84$ vs. $3.36$ on a 7-point scale, $p < .001$).
  • Social Competence: The adaptive robot also won here ($M = 4.85$ vs. $3.53$, $p = .019$).
  • Embarrassment: While the adaptive group reported lower average embarrassment ($M = 2.75$ vs. $3.85$), the result was statistically inconclusive ($p = .130$).
  • Manipulation Checks: 100% of the adaptive group noticed the robot "looking around," and 60% explicitly believed the robot had noticed the bystander.

Limitations and Future Work

  • Sample Constraints: The study used a relatively small, non-randomized sample from a single university community.
  • Static Bystanders: The bystander was stationary and didn't interact. Real-world scenarios involve moving people and varying levels of "familiarity" (e.g., a friend vs. a stranger), which the authors found might influence embarrassment.
  • Feedback Type: The feedback was intentionally "snarky" and performance-independent. Future work should explore how these behaviors affect learning outcomes when using constructive, performance-based feedback.
  • The "Why" Factor: Future studies could test if the robot explicitly explaining its privacy-seeking behavior (e.g., "I'm showing this on screen because someone else is here") improves or hinders the user experience compared to purely non-verbal cues.

Summary updated

Abstract

Robots in shared settings may expose information intended for an individual user to nearby bystanders. Robot tutoring makes this concern particularly salient: critical feedback is directed at an individual learner, yet bystanders may overhear or see it. To address this, we compared context-invariant and context-adaptive robot behaviors in a between-subjects study with 40 participants. Following an initial phase without a bystander, a bystander entered the setting. Participants then received two scripted critical feedback messages concerning their pronunciation and handwriting, independent of their actual performance. With the context-invariant behavior, the robot tutor delivered both messages aloud. With the context-adaptive behavior, the robot tutor scanned for the bystander, turned its body and display away after detecting them, gave a spoken attention cue, and presented the same feedback messages silently on its display. Participants rated the robot tutor higher in perceived privacy awareness and perceived social competence in the context-adaptive condition than in the context-invariant condition, while evidence for reduced experienced embarrassment was inconclusive. These findings inform the design and evaluation of robots that respond to bystander presence by adapting how they communicate potentially sensitive information.
BibTeX
@misc{nilgar2026you,
  author = {Atikkhan Faridkhan Nilgar and Manuel Dietrich and Kristof Van Laerhoven},
  title = {{For You, Not for Everyone: Bystander-Aware Delivery of Critical Feedback in Robot Tutoring}},
  year = {2026},
  eprint = {2611.05001},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05001},
}
Figure