A recent study from the Oxford RobotCycle project measured exactly where a cyclist looks while navigating bike lanes, bus lanes, and roundabouts through Oxford, UK. The results show that the width and separation of a lane directly affect a cyclist's cognitive load — and that most bike lanes fail to provide the same sense of safety as a bus lane.
What the Researchers Built
Researchers from the University of Oxford’s RobotCycle project equipped a cyclist with a mobile eye-tracker and rode through the city of Oxford over several months in 2024 and 2025. The routes included bike lanes, bus lanes, shared paths, roundabouts, signalised crossings, unsignalised turns, and zebra crossings. The goal was to capture high-resolution eye gaze data — fixation duration, horizontal gaze dispersion, and fixation frequency — across real-world traffic conditions.
The dataset is part of a larger effort to understand how cyclists perceive risk in urban environments. Unlike lab-based studies, this work captures natural riding behaviour with actual car, bus, and pedestrian traffic. The researchers annotated every second of video with infrastructure type, event type, and weather conditions. The result is a rich, open dataset that can help city planners, vehicle manufacturers, and autonomous driving systems understand what cyclists actually look at — and what makes them feel safe or stressed.
Key Results
The most striking finding is that bike lanes in Oxford provide almost no measurable safety benefit over riding in a car lane. For fixation duration and gaze dispersion, the differences between bike lanes and car lanes were tiny: dispersion was only 8.5% larger in bike lanes (p < 0.001, Hedges’ g = 0.09, ratio = 1.085) — a statistically significant but practically small difference.
The real contrast came with bus lanes. Cyclists in bus lanes showed longer fixation durations and much lower fixation frequencies than in bike lanes or car lanes. The table below summarises the fixation frequency data:
| Infrastructure Type | Fixation Frequency (fixations per minute) | Interpretation |
|---|---|---|
| Bike lane | 142.2 | Rapid, frequent re-fixations — high cognitive load |
| Car lane | 80.5 | Moderate scanning |
| Bus lane | 62.8 | Long, steady fixations — lower cognitive load |
Bike lanes had more than double the fixation rate of bus lanes. The researchers attribute this to Oxford’s typical bike lanes: narrow, unpainted protection, and barely separating cyclists from cars. Bus lanes, by contrast, are wide, often separated by solid white lines, and local bus drivers are accustomed to sharing the road.
Fixation duration alone did not reliably discriminate between different event types like roundabouts, crossings, or turns. This suggests that where a cyclist looks matters more than how long they hold a gaze.
How It Works
The study used a mobile eye-tracking system (likely head-mounted cameras and infrared sensors) to record the cyclist’s eye movements at 60 Hz. The raw data was processed to extract fixations — periods where the gaze remains relatively stable on a point — and calculate two key metrics:
- Fixation duration: How long the eye lingers on a single point. Longer fixations typically indicate focused attention or cognitive processing.
- Horizontal gaze dispersion: The spread of gaze along the horizontal plane. Higher dispersion suggests more scanning of the environment, often due to perceived threat or uncertainty.
The researchers log-transformed the data to correct for right skew, then used Welch’s ANOVA for omnibus testing followed by Games-Howell post-hoc comparisons. Effect sizes were reported using Hedges’ g and geometric mean ratios.
An important nuance: The dataset labels infrastructure based on where the bike actually is, not what is available. For example, if a bike lane exists but the cyclist is in a turn lane, the segment is labelled as “car lane”. This reflects real behaviour — cyclists often leave protected lanes to make turns or avoid obstacles.
Why This Matters for Robotics
For anyone building autonomous vehicles or shared-space robots, understanding cyclist behaviour is critical. The paper’s findings have direct implications:
- Perception systems: Autonomous cars should expect that cyclists in narrow bike lanes exhibit high-frequency, rapid eye movements — a signal of alertness and potential unpredictability. In contrast, a cyclist in a wide bus lane may be more relaxed, but also less reactive.
- Navigation planning: If a robot or autonomous vehicle wants to predict cyclist intent, gaze metrics could be a useful proxy for stress or upcoming manoeuvres. The data suggests that a cyclist fixating on a particular object (e.g., a car door or a turning vehicle) may be about to swerve.
- Infrastructure design: City planners and fleet operators considering warehouse robots or delivery bots that share roads with cyclists can use these insights to design safer shared spaces.
The broader RobotCycle dataset aims to eventually enable gaze-based “risk maps” of cities, which could be used by autonomous shuttles or humanoid robots navigating crowded urban environments.
Limitations and Open Questions
The study has several notable limitations. First, it involves only one cyclist — a research team member — and while routes were diverse, individual rider behaviour may not generalise. Second, the analysis was restricted to two metrics (fixation duration and gaze dispersion); adding inertial measurement unit (IMU) data to track head movements could improve interpretation.
Most importantly, the researchers could not determine what the cyclist was looking at during each fixation. Knowing that a cyclist stares at a car door rather than the road ahead would be far more predictive of stress. The team plans future work that projects gaze onto 3D scene points to answer exactly that question.
Finally, the bike lanes in Oxford are particularly narrow and unprotected. Results may differ in cities with separated cycle tracks or wider lanes. The study should be replicated across multiple cities and cyclists before drawing strong policy conclusions.
Frequently Asked Questions
Do bike lanes really not help cyclists feel safer? In this study, the bike lanes were so narrow and poorly separated from car traffic that the cyclist’s gaze behaviour was nearly identical to riding in a car lane. Wide, well-protected bus lanes showed a clear benefit.
What does fixation frequency tell us about cognitive load? Higher fixation frequency (more eye movements per minute) suggests the cyclist is scanning the environment more rapidly, often due to perceived danger or uncertainty. This is associated with higher mental workload.
Can this data be used to train autonomous vehicles? Yes. The eye gaze patterns provide a ground truth for what a cyclist considers salient. Future autonomous systems could use similar gaze models to predict cyclist behaviour in real time.
Was the cyclist’s safety compromised during the study? The cyclist was an experienced rider and followed all traffic laws. The study recorded natural behaviour in real traffic, but no risky interventions were made.
Conclusion
The Oxford RobotCycle eye-tracking study reveals that not all cycling infrastructure is equal. Wide, protected lanes like bus lanes allow cyclists to ride with longer, more relaxed gazes, while narrow bike lanes demand constant rapid scanning. These findings are a wake-up call for city planners and a valuable dataset for anyone developing robotics or autonomous systems that share the road with cyclists.