本文由 AI 分析生成
建立時間： 2026-07-30 來源： https://arxiv.org/html/2603.00154v1
Summary
This pilot study (n=29) investigates how affect-responsive robot interaction policies influence user trust during fully autonomous in-person collaboration using a Misty-II robot. Participants in the responsive condition (robot adapts based on inferred interaction state) showed ~26-point higher TI-HRC scores and ~15-point higher TPS-HRI scores compared to a neutral/reactive control condition, despite no significant difference in task accuracy. The study identifies communication viability—speech recognition reliability—as a prerequisite for the trust benefit to manifest.
本研究以 Misty-II 機器人進行全自主 HRC 實驗，比較「情感回應式」與「中性反應式」互動政策對信任感的影響。回應式條件的參與者在 TI-HRC 量表高出約 26 分、TPS-HRI 高出約 15 分。語音辨識的穩定性被識別為信任效益出現的前提條件，而非僅是調節變數。
Prerequisites
- Trust in Human-Robot Interaction (HRI) — Understanding the TPS-HRI and TI-HRC measurement scales is required to interpret the quantitative results, as both operationalize trust in HRC contexts.
- Affect inference in dialogue systems — The robot’s responsive policy depends on inferring participant emotional/engagement states from speech and conversational context; understanding this pipeline is necessary to evaluate the experimental manipulation.
- Grounding and repair in spoken dialogue — Communication viability is the paper’s key moderating construct; familiarity with Clark’s grounding theory explains why ASR failures collapse the responsive policy’s advantages.
Core Idea
The key insight is that trust in HRC reflects interaction process quality, not task outcome quality. By detecting engagement cues and deploying encouragement (36% of responsive turns) and empathy (13% of responsive turns), the responsive robot creates a higher-quality interaction process that generates trust independent of whether the collaborative task succeeds. However, this process-based trust building is entirely contingent on the communication channel remaining viable: when ASR failure rates accumulate, the robot cannot infer interaction state accurately, its responsive behaviors become inappropriate, and the trust advantage disappears. This suggests that affect-responsiveness and robust speech processing must be developed jointly rather than treating language understanding as a separable substrate.
Results
| Measure | Responsive | Neutral/Reactive | Delta |
|---|---|---|---|
| TI-HRC (post-interaction) | higher | lower | ~+26 pts |
| TPS-HRI (post-interaction) | higher | lower | ~+15 pts |
| Task accuracy | similar | similar | n.s. |
| Encouragement rate | 36% | 0% | +36 pp |
| Empathy rate | 13% | 0% | +13 pp |
| Communication breakdown rate | 25% | 22% | n.s. |
Note: 5 of 29 sessions excluded due to severe ASR failure; primary analyses on n=24 eligible sample.
Limitations
- Author-stated: Pilot sample size limits statistical power; substantial variability in participants’ functional spoken-language proficiency; no non-embodied comparison condition; affect inference relied primarily on speech and conversational context.
- Unstated: Single robot platform (Misty-II) limits generalizability; both tasks were puzzle/reasoning-based rather than physical manipulation HRC; no longitudinal measurement of trust stability; the interaction policies were researcher-designed, not learned—generalizability of findings to learned policies is unknown.
Reproducibility
- Code: Not available
- Datasets: Custom experimental data; not released
- Compute: No GPU requirements; Misty-II platform required for replication
Insights
The framing of communication viability as a prerequisite rather than a moderator is a meaningful conceptual contribution: it implies that affect-responsiveness cannot simply be “added on top” of a baseline system that has mediocre speech recognition. Systems must achieve robust grounding first. The finding also challenges purely performance-based trust models—trust in robots may be more analogous to interpersonal trust, where relational quality matters independently of competence demonstrations.
Connections
- VLA models — VLAs used as robot controllers face the same viability problem: multimodal input failures cascade into interaction quality failures
- HRI — Directly in the domain; extends HRI trust literature to fully autonomous systems rather than Wizard-of-Oz setups
- Learning from Demonstration — Trust is a prerequisite for users to provide high-quality demonstrations; this paper quantifies how interaction policy affects that trust
- dialogue-management
- social-robotics
Raw Excerpt
“Responsive interaction was associated with significantly higher post-interaction trust under viable communication conditions, despite no reliable differences in overall task accuracy… Trust advantages were most pronounced under viable interaction conditions and attenuated when communication breakdown accumulated.”