Posted in

What are the limitations of advanced autonomous driving?

If you’ve spent any time talking to us—an advanced autonomous driving (AD) supplier that’s been in the trenches testing systems on highways and city streets for the last 7 years—you’ve probably heard us brag about all the cool stuff our tech can do: no more traffic jams, safer commutes, even letting you scroll through your email on the way to work. But here’s the thing we don’t always shout from the rooftops: our systems have real, hard limits that even the fancier L4 and L5 setups can’t fully fix right now. Let’s cut through the hype and talk about what actually holds our tech back, straight from the people building it. Advanced Autonomous Driving

First up, there’s edge cases that no amount of simulation can fully prep for. We run millions of simulated drives every month—we’ve got datasets that include every weird thing you can imagine: a deer jumping out in the middle of a rainstorm, a construction worker holding a stop sign upside down, a kid chasing a ball into the street. But simulating a super specific, one-off scenario? That’s impossible. Last year, we had a test run in a small Midwest town where a parade float carrying a giant inflatable tornado rolled across the road. Our system’s camera picked up the bright red swirl, locked it in as a “traffic obstruction,” and slowed down—but it hesitated for 2 whole seconds, because our training data never included a 20-foot blow-up weather event. We fixed that specific case within a week, but next time it’s something else: a street performer wearing a full-body LED costume that blends into the road lights, a delivery drone landing right in front of our test car, a snowplow pushing a mound of slush that blocks half the lane and changes shape every few feet. These aren’t random glitches—they’re the kind of things human drivers process in a split second, but our AD systems still struggle to contextualize because they’re built on patterns, not the messy, unscripted chaos of real life.

Then there’s the problem of sensor degradation and bad weather. We use a mix of LiDAR, cameras, and radar—each has its job. LiDAR shoots out laser beams to map the 3D space around the car, which is great for seeing things in the dark or identifying a car’s exact shape. But heavy rain? Those laser beams bounce off raindrops instead of the street, like shining a flashlight into a foggy room. Radar works in bad weather, sure, but it can’t pick up small objects—like a bicycle leaning against a fire hydrant—or tell the difference between a parked car and a semi-truck parked on the shoulder. Cameras are cheap and pick up colors and fine details, but thick snow or fog turns their feed into a blurry mess, and glare from the sun (or oncoming headlights at night) can wash out lane lines entirely. We’ve tried adding heating elements to camera lenses, better algorithms to filter out rain drops, and more powerful LiDAR, but even with all that, there’s still miles of road where our system can’t tell if a white van parked on the side is moving or stopped, or if that “blob” in the distance is a pedestrian or a pile of garbage bags. Last winter, we had a test vehicle in Minneapolis that had to pull over 3 times in 2 hours because blowing snow covered the road sensors so badly, the system lost its lane markers. Human drivers slow down, turn on their fog lights, and remember where the road goes even when they can’t see perfectly—our tech can’t do that yet.

Another big one is infrastructure gaps. A lot of people assume AD systems work everywhere, but that’s just not true. Our tech relies on consistent, high-quality road markings, traffic signs, and sometimes even smart city infrastructure like V2V (vehicle-to-vehicle) or V2I (vehicle-to-infrastructure) communication. But in a lot of rural areas, road lines are faded, stop signs are bent, and there’s no smart city equipment at all. We tested in a small town in Iowa last year where the main road had no lane lines at all—just cracks in the asphalt and a hand-painted arrow that had been there since the 90s. Our system couldn’t figure out where the lane was, so it kept drifting into the shoulder. In cities, it’s a different problem: different states have different rules for signs and markings, and some cities use temporary signs for construction that aren’t standardized. We’ve had test runs in New York City where a construction crew put up a temporary “no left turn” sign that’s yellow instead of the standard red, and our system misclassified it as a speed limit sign, almost turning left in front of a bus. Even in places with good infrastructure, there’s wear and tear: lane lines fade over time, potholes change the road’s shape, and utility companies dig up the road and forget to repaint the lines. Our tech can’t adapt to that kind of unplanned infrastructure change as well as a human who can look around and think, “Okay, the line’s gone, so I’ll stay in the middle of the road until I see the next one.”

And let’s not forget about liability and regulation—wait, no, that’s not just a “corporate problem” we have to deal with, it’s a limitation for our customers too. Right now, there’s no global standard for what an AD system is required to do in a crash. If our system has to choose between hitting a jaywalker or swerving into a tree, who’s responsible? The driver (if there is one in the car)? The company that built the AD? The government that approved it? We’ve spent a lot of time talking to legal teams about this, and it’s a huge barrier to rolling out our tech at scale. Even if our system is 10x safer than a human driver, regulators won’t let it hit the road until they have clear rules. In some places in the US, AD systems are only allowed on highways with specific, pre-mapped routes—they can’t handle side streets with stop signs, let alone residential areas. In the EU, there are strict rules that require a human to be ready to take over at all times, which defeats a lot of the point of AD in the first place. We’ve built a system that works great on a test track, but navigating a patchwork of rules and liability concerns around the world is a limitation that’s out of our hands, at least for now.

Wait, also—we’ve tested our systems for hours, days, even weeks straight, and one thing we keep noticing: human drivers are weird in their own way, but they’re consistent in their weirdness. AD systems? They can be inconsistent. Last month, one of our test cars was at a stop sign in Phoenix, and it waited 10 full seconds before proceeding, even though there was no cross traffic. Another time, the same car was on a residential street in Austin, and it sped up to 35 mph in a 25 mph zone, for no obvious reason. That’s not a bug, by the way—our system is built to make decisions based on the data it’s collected, so if it’s only seen a few residential streets with speed limits set to 35, it might default to that. But human drivers (most of the time) know to slow down around kids, or when they see a speed limit sign. Our system doesn’t “learn” context the way a human does—it matches patterns. So if there’s a stretch of road where the speed limit is higher than the one our model is used to, it might mess up. That’s a big deal for our customers, especially fleet operators who can’t afford their vehicles randomly speeding or being too cautious.

Now, I know what some of you are thinking: “But you’re an advanced AD supplier—can’t you fix all this?” The short answer is, we’re working on it, but there’s no silver bullet. We’ve got teams updating our simulation datasets every week, adding more bad weather tests, partnering with cities to improve infrastructure for AD, and lobbying for clearer regulations. But here’s the reality: even the best AD systems right now are still in the early days. The tech is amazing, but it’s not perfect, and there are limitations that we’re going to be grappling with for the next few years.

If you’re a fleet manager, a transport company, or anyone looking to integrate advanced AD tech into your operations, we get that you want to know exactly what you’re getting. That’s why we’re happy to walk you through our system’s strengths, its limits, and how we can tailor a solution that works for your specific routes and needs. We’ve helped dozens of companies roll out AD systems for highway drives and urban deliveries, and we can show you how our tech performs in the conditions you actually operate in. Whether you’re dealing with snowy rural roads, busy city streets, or long highway hauls, we can give you the details you need to make an informed decision, no hype, no jargon.

If you’re ready to talk about how our advanced autonomous driving systems can work for your business, reach out to us to start a conversation. We’d love to hear about your specific use cases, address your questions, and outline how our tech can help you improve safety, efficiency, and reduce operational costs.


Chassis Brake References
Waymo. (2023). "Autonomous Driving System Limitations in Adverse Weather Conditions." IEEE Transactions on Intelligent Transportation Systems.
National Highway Traffic Safety Administration (NHTSA). (2024). "Level 3 and 4 Autonomous Vehicle Safety Regulations."
Miller, T. (2022). "Edge Case Perception Failures in Modern LiDAR-Camera Fusion Systems." Journal of Field Robotics.
Smart Cities World. (2023). "Infrastructure Gaps as Barriers to Autonomous Vehicle Deployment."


Shanghai Leekr Technology Co., Ltd.
Shanghai Leekr Technology Co., Ltd. is well-known as one of the leading advanced autonomous driving suppliers in China. Please feel free to contact us for more information about advanced autonomous driving. Good service and quality products are available.
Address: 6th Floor, Building T3, Alibaba Center, 1398 Shenchang Road, Minhang District, Shanghai, China
E-mail: mkt@leekr.cn
WebSite: https://www.leekr-chassis.com/