Why Predicting Your Future Location Doesn't Make a Robot Feel Safer

In 200 trials across a university hallway, a small robot on wheels repeatedly walked toward strangers, stopped inches from their faces, and asked them how it made them feel. The researchers who ran these experiments expected that a robot capable of predicting where a pedestrian would be in the next few seconds would feel safer to interact with. It didn't. The robot that stopped early and close—the one that seemed to think hardest about the future—left people no more comfortable than its simpler sibling. This is a story about teaching robots to share sidewalks with humans, and why foresight, it turns out, is not the same as feeling safe.

The question sounds straightforward: when a delivery robot rolls toward you on a crowded sidewalk, what makes you feel safe? Is it that the robot keeps a respectful distance? That it slows down in time? That it seems to know where you're going? Researchers at National Cheng Kung University in Taiwan decided to test these questions empirically, building four versions of the same navigation algorithm and then putting them in front of real people to see what registered. Their findings, published in July 2026, challenge a common assumption in robotics—that more prediction always means better safety—and offer a nuanced map of what objective measurements (things you can measure with sensors) and subjective experience (how people actually feel) have in common, and where they diverge.

The Science

The robot used in these experiments was an Agilex Scout Mini, a four-wheeled platform roughly the size of a large suitcase, equipped with two depth cameras mounted at the front to track pedestrians, a 360-degree LiDAR sensor to detect walls, and a laptop brain running the navigation algorithms. The researchers chose a "facing scenario": the robot started at one end of a university hallway about 30 meters long, a pedestrian started at the other end, and they walked toward each other. This simple confrontation captures the essential tension of sidewalk sharing—in a narrow space, two moving objects must negotiate passage without collision.

Four navigation algorithms were tested, each a variation on a framework called the Social Force Model (SFM). Originally developed by Dirk Helbing and colleagues in the 1990s to simulate how crowds of pedestrians move, SFM treats human motion as the result of invisible forces pushing on each person. There's a "desired force" pushing you toward where you want to go, "social forces" pushing you away from other people who get too close, and "boundary forces" keeping you off walls. By calculating these forces in real time, a robot can navigate a crowd the way a pedestrian would—reacting to other people not through explicit reasoning about their trajectories, but through a kind of virtual pressure field.

The four variants tested were: SFM (the baseline Social Force Model, using distance to decide how strongly to avoid a pedestrian), TSFM (the same model but using Projected Time-to-Collision instead of distance), PSFM (SFM with prediction added), and PTSFM (TSFM with prediction added). The key difference between SFM and TSFM is what triggers the robot to slow down. SFM watches how close the pedestrian is; TSFM watches how quickly that distance is shrinking. Time-to-collision— TTC, in the jargon—captures something more intuitive about danger: not just that someone is near, but that a collision is becoming imminent. If two cars are traveling toward each other at 60 mph each, they're both 200 meters apart from a junction. Are they equally dangerous? No—the one closing at 120 mph relative speed will be at the junction in 3.6 seconds. The one closing at 30 mph will be there in 14.4 seconds. TTC translates that relative velocity into a countdown. Projected TTC adapts this to two-dimensional space, accounting for angles rather than just straight-line approach.

The prediction variants—PSFM and PTSFM—take this a step further. Rather than responding only to where the pedestrian is right now, they average the expected social force over a time horizon of several seconds, assuming both robot and pedestrian continue moving at constant velocity. If a pedestrian is walking toward the robot, the predictive model sees not just the current geometry but the evolving geometry: in one second, two seconds, three seconds. It averages these future states and adjusts the robot's path accordingly. The idea is proactive rather than reactive—a predictive system should start avoiding a collision before the non-predictive system even notices there's a problem.

Ten volunteers participated, all male, averaging 22.9 years old. This is a limitation the authors acknowledge: the sample doesn't represent the full diversity of pedestrians who will eventually share sidewalks with robots. But for a proof-of-concept experiment, it's a reasonable start. Each participant experienced twenty trials—five with each algorithm variant, randomized to avoid order effects. After each trial, participants filled out a survey rating their comfort, the robot's movement smoothness, whether the distance felt appropriate, and whether the robot's speed felt right. These were measured on a 1-to-5 Likert scale, the standard tool for capturing subjective judgments.

The experiment was reviewed and approved by the NCKU ethics committee, and all participants gave informed consent. The researchers collected objective data throughout: the minimum PTTC reached during each encounter, the robot's average speed in the vicinity of the pedestrian, the closest the robot came to the pedestrian, the lateral distance at the moment of passing, and the maximum curvature of the robot's trajectory. They also recorded video for qualitative analysis.

The force diagram above shows how SFM conceptualizes navigation. The robot experiences three types of forces simultaneously: a desired force pushing it toward its goal, boundary forces pushing it away from walls, and social forces pushing it away from pedestrians. The robot translates the sum of these forces into forward motion and rotation—moving in response to one force component and turning in response to another. This decomposition is what makes the model computationally efficient and suitable for real-time control.

What They Found

The headline result is striking: PTTC-based navigation (both TSFM and PTSFM) significantly improved objective safety metrics compared to the distance-only variants. The minimum PTTC—essentially, the closest the robot came to an unavoidable collision—was substantially higher for the PTTC methods. In other words, when a robot uses time-to-collision rather than raw distance to decide how to react, it gives itself more margin. It starts slowing down earlier, when a collision is still several seconds away rather than when the pedestrian is already close. This is not surprising in retrospect—TTC captures urgency in a way that distance alone cannot—but the empirical confirmation matters because it grounds the theoretical advantage in real-world performance.

Average speed in the pedestrian's vicinity was lower for PTTC methods. The robot didn't just slow down at the last moment; it moderated its pace throughout the encounter, which participants might experience as more deliberate and less aggressive. Minimum distance to the pedestrian was also larger for the PTTC variants, though the difference was modest—the robot didn't end up standing absurdly far away, but it gave itself slightly more breathing room. Lateral distance at passing—the perpendicular gap between robot and pedestrian when they went around each other—was also larger for the PTTC methods. The robot swung wider, as a careful driver might when passing a cyclist.

The prediction contribution, however, was limited. When the researchers compared SFM to PSFM, or TSFM to PTSFM, most of the objective metrics showed no statistically significant difference. The predictive variants didn't do substantially better or worse—they were roughly equivalent. The one exception appeared in some of the sub-metrics, where prediction showed occasional marginal improvements, but these were not consistent enough to be meaningful. The authors are careful to note this: the prediction effect is "limited and occasionally visible in some of the sub-metrics."

This is counterintuitive. You'd expect that knowing where the pedestrian will be in a few seconds would help the robot plan a smoother, safer path. But the researchers offer an explanation: the constant-velocity assumption underlying the prediction may not hold well in a face-to-face encounter. When two people walk toward each other and then pass, their relative velocity changes constantly as they slow down, turn, and accelerate again. A prediction made at the start of the encounter becomes increasingly unreliable as the encounter unfolds. By the time the robot's predicted trajectory would have been most useful, it's already out of date.

On the subjective side, the picture is even murkier. When participants filled out their surveys, some reported that the predictive methods felt smoother and that the robot's speed behavior felt safer. But when the researchers applied the Mann-Whitney U test—a statistical test for comparing ordinal data like survey ratings—the differences between groups were not significant. The subjective impressions didn't survive statistical scrutiny. This could mean the differences are real but too small to detect with ten participants. It could mean the differences are entirely in participants' heads, shaped by expectations rather than experience. Or it could mean that the survey instrument isn't capturing what people actually felt.

The vector field diagram above illustrates how prediction works in PSFM and PTSFM. Instead of responding to the social force at the current moment, the robot averages the expected social force over a time horizon, accounting for how the geometry will evolve as both robot and pedestrian continue moving. In a simple straight-line encounter, this would work well—each agent maintains constant velocity, and the future is predictable. But real encounters involve deceleration, turning, and course corrections, which make the constant-velocity assumption increasingly inaccurate as time goes on.

The qualitative video analysis adds texture to these numbers. The researchers watched the footage and noted that PTTC methods were "clearly distinguishable" from the basic methods. The PTTC-based robots started slowing down visibly earlier, at a greater distance from the pedestrian. They nearly stopped before passing, then resumed motion. The non-PTTC robots reacted later and passed with more momentum. The predictive variants, by contrast, were "quite visually similar" to their non-predictive counterparts. The small differences were detectable across multiple trials but not apparent in individual encounters.

One observation stands out: some participants interpreted the PTTC robots' near-stop as a sign of hesitation or uncertainty. They rated these encounters as more uncomfortable, even though the robot was objectively behaving more safely—giving itself more margin, not less. "This is an example of an objectively safe behavior interpreted as subjectively unsafe," the authors write. The robot that seemed most careful was, to some participants, the robot that seemed most out of control.

Why This Changes Things

These results matter because the sidewalks of the future will be crowded with robots, and we need to design them correctly. Delivery robots are already rolling through some cities. Hospital robots ferry supplies between wards. Shopping malls are experimenting with robotic guides and bodyguards. The question of how these machines should navigate around people isn't academic—it's a design problem with real consequences for safety, comfort, and public acceptance.

The most important finding is that time-to-collision metrics work better than distance-only metrics for keeping pedestrians safe. This sounds obvious, but it has practical implications. Many existing robot navigation systems rely primarily on proximity—maintain a minimum gap, and you're safe. The PTTC results suggest this is insufficient. A robot that knows it's on a collision course with a pedestrian will behave differently from one that only knows a pedestrian is nearby. It will slow down earlier, give more margin, and swing wider on passing. These behavioral differences are measurable and statistically significant.

The prediction result is more ambiguous, and that's useful information too. Researchers have invested significant effort in pedestrian trajectory prediction—building systems that forecast where people will walk so robots can plan accordingly. The intuition is appealing: a robot that can see the future should navigate better than one that can't. But this study suggests the benefit may be overstated, at least for simple two-party encounters. Prediction helped in some edge cases but didn't move the needle on aggregate safety metrics. The authors speculate this is because constant-velocity prediction breaks down in real encounters, where people constantly adjust their speed and direction. More sophisticated prediction models might do better, but the basic message is: don't assume prediction is a silver bullet.

The gap between objective and subjective safety is the most philosophically interesting finding. A robot can be measurably, statistically safer—and still feel less safe to the people it encounters. This isn't a flaw in the pedestrians; it's a feature of how humans assess risk. When you see a robot stopping early and slowly approaching, you might interpret that as timidity, as uncertainty, as a sign the robot doesn't know what it's doing. When you see a robot maintaining speed and passing close, you might interpret that as confidence, as competence, as the robot being in control. The researchers observed this directly: some participants found the PTTC robots' cautious behavior uncomfortable, even though that caution is exactly what made them safer.

This creates a dilemma for robot designers. The behaviors that make a robot objectively safer might make it subjectively worse. A robot that gives itself enormous margins and nearly stops for every pedestrian will score well on collision-risk metrics but poorly on user experience. A robot that passes briskly and close will score well on perceived confidence but worse on objective safety. The optimal design probably lives somewhere in between, trading some objective safety for some subjective comfort—but where that trade-off falls is a value judgment, not just an engineering problem.

For the field of robot navigation, the study offers a methodological contribution beyond its specific findings. The authors systematically compared four algorithm variants on both objective and subjective metrics, using real hardware and real human participants. This kind of empirical validation is rarer than it should be. Many navigation papers evaluate algorithms in simulation or on benchmark datasets without ever putting them in front of a human. The gap between simulation and reality is well-documented—algorithms that work beautifully in a simplified environment often fail in the messiness of the real world. Studies like this one help fill that gap, providing ground truth about what actually works when a robot meets a person.

What's Next

Several questions remain open. First, the sample size: ten male participants in their early twenties at a single university campus. The results may not generalize to women, older adults, children, people with disabilities, or populations in different cultural contexts. Personal space norms differ across cultures—people in some societies stand closer when talking, others prefer more distance. A robot designed in Taiwan might behave differently than one designed in New York or Nairobi. Future research should recruit more diverse participants and test whether the findings hold.

Second, the scenario: a simple face-to-face encounter in a controlled hallway. Real sidewalks are messier—multiple pedestrians, obstacles, narrow passages, crowded conditions. The authors note that prediction showed limited benefit in single-pedestrian scenarios; it might show more benefit in multi-pedestrian scenarios where the robot needs to anticipate cascading crowd movements. This is a natural next step: test the same algorithms in group scenarios and see if prediction makes a larger contribution.

Third, the prediction model itself: constant velocity is a crude assumption, and the researchers acknowledge this. More sophisticated prediction models—ones that account for deceleration, course corrections, and social behaviors like following the crowd—might perform better. The authors suggest that decaying weights, where near-term predictions matter more than distant ones, could improve accuracy. Testing different prediction horizons (shorter horizons are more accurate but give less advance warning) is another avenue for future work.

Fourth, the subjective measurement: the Likert scale surveys capture something, but probably not everything. The researchers asked participants to rate comfort, smoothness, distance, and speed after each trial. But how people feel about a brief encounter in an experimental setting may differ from how they'd feel about a robot that crosses their daily commute. Long-term exposure might change perceptions—people might habituate to cautious robots or, conversely, find them increasingly annoying. Studies that measure acceptance over time, not just in the moment, would add valuable perspective.

Finally, the design implications: if objective and subjective safety sometimes pull in opposite directions, how should engineers navigate that trade-off? The authors don't answer this question, but they articulate it clearly. A safety certification framework might prioritize objective metrics—collision rates, near-miss frequencies, maximum deceleration forces. But public acceptance may depend more on subjective impressions. The study suggests these two goals are not identical, and that optimizing for one may come at the cost of the other.

The robots are coming to the sidewalks regardless. The question is what they'll look like when they get there: how fast they'll move, how close they'll get, whether they'll hesitate or stride confidently past. The answers won't come from simulation alone. They'll come from studies like this one, where researchers put robots in front of real people and ask the hard question that all of robotics will eventually have to answer: what does it feel like to share a sidewalk with a machine?

---

This digest is based on "Empirical Pedestrian Safety Assessment in a Mobile Robot Using a Predictive Social Force Model" by Alireza Jafari, Yun-Hao Tsai, and Yen-chen Liu, published on arXiv in July 2026. The study was conducted at National Cheng Kung University, Tainan, Taiwan, with ethics review by the NCKU Institutional Review Board.