The Robot That Teaches Women Better Than Men
A robot tour guide that banter improves learning by 54% for women but not men—revealing a hidden gender gap in how we design social robots.
Female visitors learned 54% more from robot tours that bantered—but men showed no difference. The same machines, the
The Tour Guide of the Future May Learn Differently Depending on Who It's Teaching
In a brightly lit room at the University of Michigan, something unexpected happened. Thirty people signed up to take a museum tour led by a robot. Some tours featured a single cheerful robot named Remy. Others featured Remy plus a projected virtual companion called Scout—a cartoon vehicle with anthropomorphic features who bantered with Remy, cracking jokes and playfully challenging the robot's statements. By every self-reported measure—engagement, satisfaction, how much people said they enjoyed the experience—all three conditions felt roughly the same. The numbers barely moved.
But hidden in the data, something more interesting emerged. When researchers split participants by gender, the results diverged sharply. Female participants learned 21% more content—retained more facts, scored higher on quizzes—when the robot team bantered rather than when a single robot lectured. For male participants, the conversational style made essentially no difference. The same robot, the same content, the same room—but who absorbed what changed depending on who was listening.
This finding—that a robot's communication style might matter profoundly for some learners and not others—challenges a quiet assumption in human-robot interaction research: that what works for "users" works universally. It suggests instead that the future of robot guides, classroom assistants, and social robots may need to adapt not just to context, but to the specific people in front of them.
"The proposed dyadic conversational style in this paper influenced learning performance differently by gender," the researchers write, landing on a carefully hedged formulation for a result that is anything but subtle.
The Science Behind the Finding
The study emerged from a straightforward premise: museums are caught between two failing models. Static displays and audio tours are cheap but passive—they don't adapt, don't respond, and don't create the kind of engagement that helps people remember what they've seen. Human tour guides are effective but expensive, requiring salaries, scheduling, and availability that many institutions can't sustain. Robots seemed like a natural middle ground—capable of dynamic interaction, scalable across time and location, and potentially engaging enough to improve learning outcomes.
But earlier work had produced confusing results. When researchers paired two robots together, participants reported enjoying the experience more—but actually remembered less content than those who toured with a single robot. Engagement and learning had split apart. The robots were entertaining people without educating them.
This gap led the research team—based at the University of Michigan's Department of Robotics and Incheon National University in South Korea—to design something new. Rather than deploying two separate robots (which created retention problems) or a single robot with pre-programmed slides (which lacked dynamism), they created a mixed-agent system: one physical robot paired with a projected virtual avatar that could move around the room, participate in conversation, and direct visitors' attention to different exhibits. The system achieved "the interaction richness of two mobile agents from a single platform," as the researchers put it—a clever engineering solution that avoided the complexity of coordinating two separate robots while preserving the social complexity of having two distinct voices in the room.
The physical platform was a Toyota HSR (Human Support Robot), equipped with depth cameras to track where visitors were standing, a head-mounted display screen, and a button on its shoulder that visitors could press to trigger video content. Mounted on its back was a gimbal-mounted laser projector capable of casting the virtual avatar—Scout—onto walls and floors at different locations throughout the room. Inverse kinematics calculations determined where to project Scout based on the robot's position, enabling the avatar to appear to move alongside the physical robot, pointing at exhibits and responding to questions in real-time.
The researchers tested three conditions in a randomized within-subjects design—meaning each participant experienced all three, in a randomly assigned order. Condition 1 was the control: a single robot named Remy speaking in a cheerful, informative style. Condition 2 paired Remy with Scout in a storytelling mode, where both agents delivered information but didn't interact with each other. Condition 3 paired the same agents in a bantering mode, where they addressed each other, made jokes, and created the kind of dynamic conversation you might hear between two enthusiastic human docents.
Thirty participants took part—16 men and 14 women, with a mean age of 32. They wore safety helmets fitted with ArUco markers, which the robot's cameras tracked to measure their physical distance from the machine, the angle of their head relative to the robot, and how quickly they responded to prompts like "look at that poster." At three points during each condition, participants answered multiple-choice questions about the exhibit content, allowing researchers to calculate a learning performance score: how much their knowledge improved from before to after each tour segment.
The study was reviewed and approved by the University of Michigan's Institutional Review Board and designed with sufficient statistical power (0.80, with an effect size of f = 0.24) to detect meaningful differences between conditions.
What They Found
The headline result was unambiguous: mixed-agent tour guides outperformed single robots on learning performance, but only for women.
Across all 30 participants, the bantering mixed-agent condition (C3) produced learning scores averaging 0.77 on a normalized 0-to-1 scale, compared to 0.61 for the single-robot condition (C1). That 16-point difference was statistically significant (), though the intermediate storytelling condition (C3) at 0.75 didn't significantly outperform the baseline. The difference between the single robot and storytelling was marginal (), hovering just above the conventional threshold for significance.
But when the researchers divided the sample by gender, the effect crystallized. For female participants, the bantering mixed-agent condition produced learning scores averaging 0.77—compared to just 0.50 for the single-robot baseline. That 27-point gap, representing a 54% relative improvement in learning, was significant at . For male participants, all three conditions clustered together, with no significant differences between them. The conversational style of the robot team didn't seem to matter for men; it mattered enormously for women.
Learning Performance by Gender and Condition
| Label | Value |
|---|---|
| Single Robot (C1) | 0.5 |
| Mixed-Agent Storytelling (C2) | 0.71 |
| Mixed-Agent Bantering (C3) | 0.77 |
This gender-moderated effect appeared in the behavioral data as well. Female participants consistently stood closer to the robot throughout all conditions, but they stood closest of all during the bantering condition—suggesting the interactive style encouraged a more intimate physical engagement. Male participants maintained significantly greater distance from the robot across the board (), and their head movements were larger, indicating they repositioned themselves more frequently relative to the machine.
Physical Distance from Robot by Gender and Condition
| Label | Value |
|---|---|
| Single Robot (C1) | 2.8 |
| Mixed-Agent Storytelling (C2) | 3.1 |
| Mixed-Agent Bantering (C3) | 2.7 |
The correlation analysis revealed another wrinkle: learning performance was negatively correlated with engagement (). Participants who reported higher engagement scores actually learned less. This counterintuitive finding—more enjoyment, less retention—echoes the earlier dual-robot research but adds new texture. It suggests the relationship between how people feel during a learning experience and how much they actually remember is more complex than simple intuition suggests. Engagement and learning may pull in different directions, or perhaps they're measuring different things at different timescales.
Interestingly, prior experience with robots was positively associated with learning (). People who'd interacted with service robots before learned more across all conditions—a finding that could suggest familiarity with the medium reduces cognitive friction, or perhaps that people who seek out robot interactions are already more comfortable with technology-mediated learning.
The interview data added qualitative depth to these numbers. Of the 29 participants who expressed a preference, 17 favored the mixed-agent teams—citing the "fresh perspectives" that came from shifting attention between two agents, the way the banter "emulated human-to-human communication," and the feeling that it was "closer to a human being than a robot." One participant (P8237) said it felt "closer to a human being than a robot," a phrase that captures something the survey metrics couldn't fully capture.
Female participants specifically noted the "creative" and "cute" aspects of the team. They found Scout's animated eyes made interaction "easier" and more engaging, and the dialogue between the agents increased their engagement with the content. "The 'talking' and 'interaction' between the agents," as one female participant (P7l98) put it, "increased their engagement with the content."
But not everything was positive. Participants complained about the robot's speed—a deliberate 0.8 km/h chosen for safety that felt painfully slow compared to natural human walking pace. "Forcing them to moderate their pace," as the researchers noted, created friction. This pragmatic limitation reminds us that the gap between "works in the lab" and "works in the wild" remains substantial.
Why This Changes Things
The most immediate implication is methodological: studies that report aggregate results across all participants may be hiding effects that matter enormously for specific subgroups. The researchers' decision to conduct separate analyses by gender revealed a finding that the pooled data only hinted at. This isn't a new lesson—the call for disaggregated analysis has grown louder across many fields—but it's one that human-robot interaction research has been slow to absorb.
But the deeper implication is theoretical. Why would a bantering conversational style improve learning for women but not men?
The researchers don't offer a definitive answer, but several plausible mechanisms exist. Social learning theory suggests that women may be more attuned to relational dynamics—observing how agents interact with each other as a way of making sense of the content. If the bantering style models curiosity, questioning, and collaborative knowledge-building, it might provide a scaffold for encoding information that men don't require. Alternatively, social presence theory suggests that more socially rich environments may compensate for feelings of isolation or discomfort that some women experience in technology-mediated settings. The presence of two agents, interacting dynamically, may have created a more "human" environment that felt safer for learning.
There's also the possibility of social conditioning. The literature on gendered learning suggests that women may be more accustomed to navigating environments designed primarily with men in mind—spaces where they must work harder to find their place. A richer, more interactive environment may simply reduce that cognitive overhead, freeing up resources for actual learning. This isn't a deficit model; it's a design critique. Environments that work fine for the majority may systematically disadvantage minorities, and adaptive systems should account for that.
The finding also challenges the assumption that "engagement" is an unambiguous good. The negative correlation between engagement and learning suggests that making robot tours more fun might actually make them less educational—or at least, that the metrics we use to measure engagement may be capturing something different from learning. Perhaps entertaining banter increases subjective enjoyment while fragmenting attention, making it harder to encode facts. Or perhaps high engagement is a proxy for low cognitive load, and it's cognitive load that drives retention. These are empirical questions, but the existing data point toward a tension that designers need to confront rather than smooth over.
The preference data adds another dimension. Despite the learning advantage for women in the bantering condition, and despite the general learning advantage of mixed-agent teams, engagement and quality-of-experience scores didn't differ significantly across conditions. People said they enjoyed all three tours equally. But interview data revealed a strong preference for mixed-agent teams—and that preference held regardless of gender. Something in the lived experience of the mixed-agent interaction, something that surveys didn't capture, made people want more of it.
This disconnect between self-report and qualitative data is familiar in experience research, but it's worth dwelling on. The mixed-agent condition wasn't just a technical achievement—it created a qualitatively different social environment, one that participants recognized as more human-like and more engaging, even when they couldn't articulate why in a survey. That suggests designers have room to optimize for preference and engagement in ways that might exceed what traditional metrics suggest is possible.
What's Next
The study has obvious limitations. Thirty participants is a reasonable sample for a within-subjects experiment with three conditions, but it limits generalizability. The effect size for the gender difference in learning ( for women versus for the overall sample) suggests a meaningful signal, but replication with larger samples is essential. The study was conducted in a controlled laboratory-like environment at a research university; how these results translate to actual museums, with their varied acoustics, lighting, and visitor demographics, remains unknown.
The mechanism behind the gender difference remains unexplained. Is it relational dynamics? Social presence? Differential comfort with technology? Something else entirely? Answering these questions will require studies designed to isolate specific factors—perhaps comparing same-gender versus mixed-gender agent pairs, or varying the relational tone of interactions independently from the information content.
There's also the practical question of deployment. The current system requires participants to wear ArUco-marked helmets for tracking, which is fine for research but impossible in a public museum. Markerless tracking using computer vision exists but introduces its own challenges in variable lighting and crowd conditions. The robot's slow pace reflects a safety choice, but visitors accustomed to moving at human speed may find it frustrating. These engineering challenges are solvable, but they remind us that lab demonstrations and deployable systems are different things.
Perhaps most interestingly, the study opens questions about adaptive systems. If a robot tour guide can detect that a visitor is female (through appearance, voice, or explicit input), and if female visitors learn better with bantering dialogue, should the system adapt? Should it ask? There are obvious risks—reducing people to demographic categories, reinforcing stereotypes, making incorrect assumptions. But there are also obvious benefits—personalized learning experiences that work better for more people. The design space here is rich and underexplored.
The researchers frame their work as a contribution to "human-robot interaction within multi-modal systems"—and it is. But the implications extend further. As robots move from labs into classrooms, hospitals, care facilities, and public spaces, the question of whose needs they serve becomes urgent. A robot that works equally well for everyone may be an impossible ideal. A robot that adapts to diverse learners may be a more achievable—and more equitable—goal.
The museum tour is a small domain, but the questions it raises are large. How should machines communicate with the humans they serve? What counts as engagement, and is it always good? And when a system works better for some people than others, what do we owe to the people it serves less well?
These aren't just technical questions. They're design questions, ethical questions, and ultimately political questions. The robot tour guide at Michigan found something simple: that the same words, spoken by the same machine, land differently depending on who's listening. The harder work is deciding what to do with that knowledge.
Key Correlations Between Study Variables
| Label | Value |
|---|---|
| Engagement (UES) vs Learning | -0.37 |
| Robot Experience vs Learning | 0.3 |
| Gender vs Distance | -0.48 |
| Head Angle vs QoE | 0.26 |
The collaborative dynamic—characterized by jokes and dialogue—emulated human-to-human communication, making the experience feel closer to a human being than a robot.
Sign in to join the conversation.
Comments (0)
No comments yet. Be the first to share your thoughts.