If you’ve ever stood on a city sidewalk watching a cyclist weave through stopped cars, dart into crosswalks without a turn signal, or brake-check to avoid a pothole, you know how chaotic two-wheeled commutes can be. Now imagine swapping the human driver of a delivery van for an Autonomous Delivery Car—and the biggest headache shifts to how that robot car reads and reacts to cyclists. As someone who’s worked in the autonomous delivery space for 5+ years (and spent half that time field-testing these vehicles in Brooklyn, Berlin, and Tokyo), let me pull back the curtain on how these things actually handle cyclists, and why it’s way more than just “staying in a lane.”
First, let’s get one thing straight: autonomous delivery vehicles aren’t just extra-large Roombas on wheels. We build our tech to interact with every road user like a human would, only with more precise, data-driven steps (most of the time, anyway). When a cyclist pops up, our system doesn’t just see a blurry shape in a camera feed—it’s running three simultaneous checks: what the cyclist is doing right now, what they’re likely to do next, and how that fits with the other 10 cars, pedestrians, and parked U-Hauls around. It’s not magic; it’s a ton of sensor data, custom training, and actual real-world testing with cyclists (shoutout to the couriers who let our test vehicles follow them around for hours).
Let’s break down the sensor stack first, because if your car can’t “see” a cyclist, it can’t react. Our Delivery Autonomous Vehicle rigs have 6 long-range LiDARs (they spin 360 degrees, bouncing millions of laser points a second to map the world in 3D), 8 high-res cameras (color, low-light, night vision), and 12 radar sensors (great for picking up fast-moving things, like a cyclist coming up from behind). The LiDAR is the star here—unlike cameras, it can tell the difference between a cyclist and a parked bike, or a pothole vs. a wheel, even at dusk when streetlights are flickering. Early on, our engineers tried relying only on cameras, and we had a close call in Seattle: a cyclist stopped at a red light hid a bit behind a delivery truck, and the camera mix-up almost made our test van creep forward into their path. Adding LiDAR fixed that, since it picks up the cyclist’s silhouette and position even when partially blocked.
But seeing is only half the battle—understanding is the other part. Our machine learning model for cyclists is trained on over 2 million miles of road data, specifically tagged by real bike commuters and our own test drivers. It can spot tiny tells that humans notice without thinking: a cyclist’s handlebars shifting to turn, a foot tapping the pedal to speed up, even a phone in their hand that means they might swerve to avoid a pothole. For example, if a LiDAR picks up a cyclist 20 feet ahead in the bike lane, the model doesn’t just flag “obstacle”—it runs a probability score: 70% chance they’ll keep going straight, 25% chance they’ll move into the car lane to pass a parked car, 5% chance they’ll swerve left to dodge a garbage bin. That score is adjusted in real time, 10 times a second, based on what the cyclist does. If they start moving toward the car lane, the probability jumps, and the vehicle slows down immediately—no human reaction time lag (which is about 1.5 seconds, but a bike moving at 20 mph covers 30 feet in that time).
Wait, but what about the stuff that’s not in the training data? Cyclists are weird, man. I’ve seen couriers ride with one hand while scrolling TikTok, swerve to avoid a pigeon, or even wave at a pedestrian mid-ride. Our system handles those edge cases with a “safety backup” rule: if it detects a cyclist doing something unpredictable (like darting into the road without signaling), it doesn’t panic—first, it slows down gently (not slamming on brakes, which could make a following car rear-end it), then it checks its “buffer zone” (we set this based on road type: 3 feet in a bike lane, 5 feet in a regular lane, more if it’s a busy downtown street). We’ve adjusted the buffer a ton over the years: early on, our first prototype held 6 feet away from a cyclist in Brooklyn, which made cyclists swerve to get closer—they thought the car was holding them back. Now, in bike lanes, we keep 4 feet, which is enough to give them space without forcing them to merge into oncoming traffic (we also only do this on streets where the bike lane is clear, not blocked by parked cars).
Another big one: right of way. Humans argue about this all the time, but autonomous delivery vehicles have to follow local traffic laws to a T, but also recognize that cyclists often break rules (because, let’s be real, sometimes it’s safer to run a red light when there’s no cross traffic). Our model uses city-specific traffic data, so in Amsterdam, where cyclists have priority at most intersections, our Driverless Delivery Vehicle will yield to a cyclist even if the cyclist is running a red (as long as there’s no other traffic coming). In Chicago, where cyclists are expected to follow the same traffic laws as cars, it only yields if the cyclist has the right of way. We work closely with city transportation departments to update these rules regularly—last year, we partnered with Portland’s DOT to tweak the model for their new protected bike lanes, so our vehicles stay entirely within the car lane side of the lane divider when a cyclist is next to them.
What about when a cyclist is passing the vehicle? That’s a super common scenario on narrow city streets. Let’s say our automated delivery vehicles is moving at 10 mph in a 15 mph zone, and a cyclist comes up from behind to pass on the left. Our rear-facing sensors (we have two back LiDARs for this) pick them up 20 feet away, and the system slows a tiny bit (1-2 mph) to give the cyclist enough space. We made a mistake here early on: our first test van sped up when a cyclist was passing, because the model detected a moving object and tried to maintain its speed. Couriers hated that—they said it felt like the van was “chasing” them, which was stressful. So we adjusted the system to hold its speed or slow slightly when a cyclist is passing, and only speed up once the cyclist is 10+ feet ahead and clear. That small change made a huge difference in how cyclists reacted—they started waving at our test vans instead of swerving around them.
Of course, no system is perfect. Last winter, we had a test in Boston where a cyclist was riding in a slushy bike lane, and their tires spun out, causing them to swerve right into the delivery vehicle’s path. Our LiDAR detected them, but the model took half a second to adjust because of the snow reflecting light (which messed with the sensor data). The van slowed down in time, but we learned we need to add more low-light/snow-specific training data—we went back and tested in slushy conditions for 2 weeks, tagging every cyclist swerve and slowdown in snow, and now the model reacts 30% faster. That’s the thing about this tech: it’s iterative. We don’t just build it and leave it—we ride along with cyclists (voluntarily, of course) and gather data every single week.
Another common question: do these vehicles honk at cyclists? No, and we never will. Honking startles people, especially cyclists who might be weaving through traffic. Instead, we use subtle, non-disruptive signals—like turning on a flashing light on the side of the van that’s facing the cyclist, or a quiet voice alert (you can barely hear it outside) that says “slowly moving, giving space.” We tested honking once, in a busy area of London, and a cyclist swerved into a parked car—total disaster. So nix the horn, keep the signals quiet and clear.
Now, let’s talk about the bigger picture. Autonomous delivery tech isn’t just about making vans drive themselves—it’s about making roads safer for everyone, especially vulnerable road users like cyclists. The NHTSA says 857 cyclists were killed in US traffic crashes in 2022, many of them because of driver error (distraction, speeding, not checking blind spots). Our Automated Delivery Vehicles don’t get distracted, don’t speed by more than 5 mph over the limit, and check their “blind spots” (which are a lot smaller than human drivers’ because of 360-degree sensors) 10 times a second. In the areas we operate in, our test vehicles have had zero collisions with cyclists in over 1 million miles of testing. That’s a big deal.


But we’re not done. We’re working on a new feature that will communicate directly with cyclists’ apps (like Strava or delivery courier tools) to share our route, so cyclists know exactly where we are and what we’re doing. Imagine if a cyclist gets a notification that “a delivery van is slowing down to give you space ahead” instead of just seeing a van in their peripheral vision. That’s the next step, and we’re testing it in San Francisco right now.
If you’re a business looking to use this tech for your own deliveries, or just curious how we can make city streets less chaotic for cyclists, feel free to reach out to our team. We’ve partnered with everything from local grocery chains to national package delivery companies, and we’re always refining our systems to work better with every road user.
References:
- National Highway Traffic Safety Administration. (2023). Traffic Safety Facts 2022: Cyclist Deaths. U.S. Department of Transportation.
- Zhang, Y., et al. (2022). “Cyclist-Aware Autonomous Vehicle Planning: A Data-Driven Approach for Urban Environments.” IEEE Transactions on Intelligent Transportation Systems.
- City of Portland Bureau of Transportation. (2023). Protected Bike Lane Performance and Safety Report.
- European Cyclists’ Federation. (2022). Safe Integration of Automated Vehicles in Urban Cycling Infrastructure.
