💻 computer science

When May I Help You? On The Effect of Proactivity on Group Human-Robot Collaboration

This paper investigates the impact of robot proactivity versus reactivity in group human-robot collaboration within an escape room setting, revealing that while reactive models yield higher success rates, the effectiveness of either approach is significantly shaped by users' prior experience, personality traits, and group dynamics; notably, although Introverted participants were faster with the reactive robot, they actually preferred the proactive robot and rated it as more competent.

Original authors: Thomas Vitry, Vanessa Maeder, Kieran Edgeworth, Asihati Hazaiti, Doga Deniz Ates, Connor Gäde, Jan-Gerrit Habekost, Dennis Becker, Stefan Wermter

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Thomas Vitry, Vanessa Maeder, Kieran Edgeworth, Asihati Hazaiti, Doga Deniz Ates, Connor Gäde, Jan-Gerrit Habekost, Dennis Becker, Stefan Wermter

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are playing a cooperative escape room game with two friends. Suddenly, a robot joins your team. The big question the researchers asked is: How should this robot behave?

Should it act like a silent waiter who only speaks when you snap your fingers and ask for something (Reactive)? Or should it act like an enthusiastic co-pilot who constantly chimes in with ideas, asks how you're doing, and offers help without being asked (Proactive)?

The researchers set up a study where pairs of people tried to solve puzzles in a room with a humanoid robot named NICO. They tested both styles to see which one helped the team win and how it made people feel.

Here is the breakdown of what they found, using some simple analogies:

The Two Styles of Robot Behavior

The Reactive Robot (The "Waiter"): This robot sits quietly. It only talks if you say its name or ask a direct question. It's like a helpful butler who waits for your command.
The Proactive Robot (The "Co-pilot"): This robot listens to everything. If the humans stop talking for 90 seconds, it jumps in to ask, "Hey, how's it going? Need help?" It tries to keep the momentum going on its own.

The Big Results: Quantity vs. Success

More Talking: The Proactive robot definitely talked more. It was like a chatterbox that kept the conversation flowing. The Reactive robot only spoke when spoken to.
Winning the Game: Surprisingly, the teams using the Reactive robot won slightly more often (about 93% success) compared to the Proactive robot (about 71% success).
- The Analogy: Imagine driving a car. The Proactive robot is like a passenger who keeps giving you directions even when you know the way, sometimes distracting you. The Reactive robot is like a passenger who only speaks when you ask, "Where do we turn?" This allowed the humans to stay in the "driver's seat" and focus better, leading to more wins.

The "It Depends" Factor: Who You Are Matters Most

The most interesting part of the study is that there is no "one size fits all" answer. The best robot style depends entirely on the personality and experience of the humans.

Think of it like crossing a river: An expert swimmer can cross easily on their own and might find floaters cumbersome and distracting, while a beginner swimmer needs those floaters to stay afloat and reach the other side safely. Similarly, the robot's behavior needs to match the human's level of competence.

1. The "Tech-Savvy" vs. The "Newbie"

People who know Large Language Models (LLMs): These are people used to chatting with AI. They solved the early puzzles faster with the Reactive robot.
- Why? They knew exactly how to talk to the AI to get the right answers. When the AI started interrupting on its own (Proactive), it got in the way of their efficient "prompting" style.
People new to AI: They didn't have this specific skill, so the results were different for them.

2. The "Escape Room Veteran" vs. The "First-Timer"

Escape Room Veterans: People who have played escape rooms before preferred the Reactive robot. They rated it as more "likable."
- Why? They are used to solving puzzles themselves. They didn't want a robot bossing them around; they wanted a tool they could use when needed.
First-Timers: People who had never played before actually solved the final puzzle faster with the Proactive robot.
- Why? They were lost and needed a guide. The Proactive robot acted like a helpful tour guide, nudging them forward when they were stuck.

3. The "Introvert" vs. The "Extrovert"

Introverts: They solved puzzles faster with the Reactive robot. However, contrary to what one might expect, they actually preferred the Proactive robot and rated it as more competent.
- The Twist: While the quiet Reactive robot allowed them to work more efficiently without interruption, the Proactive robot’s willingness to step up and lead made them perceive the robot as more capable and intelligent.
Extroverts: They seemed to prefer the Reactive robot as well in terms of "likability," but the data showed they felt the Reactive robot was more "intelligent."
- The Insight: The study suggests that people who are comfortable in the situation (like extroverts or experts) generally prefer a robot that follows their lead (Reactive). However, for introverts, there is a nuance: while they work faster with a passive partner, they respect and prefer the competence of an active one.

The "Uncanny Valley" of Behavior

The paper mentions a concept called the "Uncanny Valley," which usually refers to robots looking too human but not quite right, making us feel creeped out.

The researchers found a similar effect with behavior:

Experienced people saw the Proactive robot as trying too hard to be human (like a nervous actor), which made them feel slightly uncomfortable. They preferred the Reactive robot because it felt more like a reliable tool.
Inexperienced people saw the Proactive robot as more "alive" and "capable." Because they didn't know the robot's limitations, they appreciated the robot taking the lead.

The Bottom Line

The paper concludes that you can't just program a robot to be "helpful" in a generic way. Helpfulness is in the eye of the beholder.

If your team is expert and confident, they want a robot that stays quiet and waits to be asked (Reactive).
If your team is nervous or inexperienced, they want a robot that speaks up, offers ideas, and guides them (Proactive).

The robot's "personality" needs to be a mirror that reflects the group's current needs. If the group is comfortable, they want a passive partner. If the group is struggling, they want an active leader.

Technical Summary: When May I Help You? On The Effect of Proactivity on Group Human-Robot Collaboration

Problem Statement

In multi-party human-robot collaboration (HRI), robot initiative presents a central design challenge. A robot that acts proactively may provide timely support but risks disrupting coordination, dividing attention, or interrupting turn-taking. Conversely, a reactive robot that waits to be addressed preserves human control but may miss opportunities to assist. This tension is particularly acute for language-enabled robots, where contributions occur through both physical action and dialogue, making them visible as both task support and social behavior. The paper investigates how a robot should participate in multi-party collaboration: by adhering to a reactive style (waiting to be addressed) or a proactive style (contributing autonomously and re-initiating interaction).

Methodology

The study employs a collaborative escape room scenario involving pairs of human participants and a humanoid robot (NICO).

Experimental Setup:
- Participants: 56 participants (28 pairs) were recruited. Condition assignment was performed at the pair level, resulting in 14 pairs in the proactive condition and 14 in the reactive condition.
- Robot Platform: The NICO robot, equipped with an expressive LED face and anthropomorphic hands, was embedded with a ChatGPT 4o Large Language Model (LLM) via the ELMiRA framework.
- Sensing and Perception: The robot utilized individual wearable microphones to capture speech, transcribed via a Whisper model. It maintained a memory system (dictionary-based) for environmental facts and could query its memory or capture images to inform the LLM.
- Task: Participants acted as interns tasked with finding a password to unlock a computer within a 60-minute window. The task involved a linear sequence of four puzzles requiring environmental observation, pattern matching, and information retrieval.
Interaction Models:
- Reactive Model: The robot functioned similarly to a speech assistant, responding only when explicitly addressed by name or when transmitting automated hints.
- Proactive Model: The robot listened continuously to all speech in the room. If no interaction occurred for 90 seconds, it autonomously re-initiated contact to ask about progress and offer help.
Measures:
- Objective Metrics: Puzzle-solving times, hint usage, and interaction frequency.
- Subjective Metrics: Pre- and post-experiment questionnaires assessing demographics, personality (BFI-10 for extraversion), and prior experience (escape rooms, LLMs, robots).
- Perception Scales: The Godspeed Questionnaire Series (anthropomorphism, animacy, likability, intelligence, safety) and the Robotic Social Attributes Scale (RoSAS: warmth, competence, discomfort).

Key Contributions

Empirical Comparison of Initiative Models: The study provides a direct comparison of proactive and reactive interaction styles in a multi-party, cooperative HRI setting, moving beyond dyadic interactions.
Contextualized Escape Room Framework: It introduces a specific collaborative escape room setup designed to examine robot initiative, featuring a puzzle graph that tracks progress and allows for controlled hint delivery.
Moderating Variables Analysis: The research explicitly investigates how user prior experience (with escape rooms, LLMs, and robots) and personality traits (extraversion) moderate the effects of robot initiative on performance and perception.

Results

Interaction Frequency: The proactive model significantly increased the number of interactions ( $p < .001$ ) compared to the reactive model (Mean 97.86 vs. 29.07).
Task Success: The reactive model showed a descriptively higher overall success rate (92.86%) compared to the proactive model (71.42%), though this difference was not statistically significant ( $p = .077$ ).
Variance in Perception: While mean scores for perception scales (Godspeed and RoSAS) did not differ significantly between models, the proactive model exhibited significantly lower variance in Animacy and Likeability (Godspeed) compared to the reactive model.
Moderating Effects of Experience and Personality:
- LLM Experience: Participants with LLM experience solved early puzzles (1 and 2) significantly faster with the reactive model. However, for Puzzle 4, they were faster with the proactive model.
- Robot Experience: Participants with prior robot experience solved Puzzle 3 significantly faster with the proactive model. In terms of perception, experienced users rated the reactive robot higher in Intelligence, while inexperienced users rated the proactive robot higher in Competence and Animacy.
- Extraversion: Introverted participants solved Puzzle 2 faster with the reactive model but Puzzle 4 faster with the proactive model. Despite being faster with the reactive model in Puzzle 2, introverts preferred the proactive model and rated it higher in Competence, whereas extroverts rated the reactive model higher in Intelligence and Likeability.
- Escape Room Experience: Novices solved Puzzle 4 faster with the proactive model and rated it higher in Animacy and Competence. Experienced escape room participants rated the reactive model higher in Likeability.

Significance and Claims

The paper claims that the effects of robot initiative are not universal but are simultaneously shaped by users' prior experience, personality traits, and the specific needs of the group.

Heterogeneity of Preference: The study demonstrates a "heterogeneity preference" where users uncomfortable with the situation (novices, introverts) tend to favor the proactive robot, perceiving it as more competent and human-like (standing "above" the uncanny valley). Conversely, users comfortable with the situation (experienced, extroverts) tend to favor the reactive robot, perceiving it as more intelligent and less intrusive (falling "within" the uncanny valley).
LLM Interaction Patterns: The findings suggest that LLM experience leads to more efficient interactions, but only when the user fully controls the interaction (reactive model). Proactive robots, which intercept inter-human communication, may disrupt these established patterns, complicating the robot's acceptance as an in-group social partner.
Design Implications: The authors conclude that designing robots for social settings is complicated because users bring different expectations based on their familiarity with the technology and the task. The robot's social image is tied to the role it is expected to fill, implying that a single interaction model may not suffice for diverse user populations. The paper suggests that future work should explore specialized LLMs designed explicitly for embodied HRI to better align with natural human-human communication patterns.