Dead Bugs Are Getting in the Way of Fully Autonomous Self-Driving Cars

Photo: shanecotee/Getty Images

On October 19, 2016, Elon Musk kicked off the lie about Tesla Full Self-Driving capabilities: “The basic news is that all Tesla vehicles exiting the factory have the hardware necessary for Level 5 autonomy. So that in terms of the cameras and compute power, it’s every car we make. So, on the order of 2,000 cars a week are shipping now with Level 5, meaning hardware capable of a full self-driving or driverless capability.”

Don’t get me wrong—I am not among those in the $TSLAQ community on Twitter who believe that everything Musk touches is a massive fraud. While there is much about the accounting around his businesses that is highly questionable and very possibly fraudulent, I think Musk is a true believer in most of the ideas he has brought forth, regardless of how outlandish many seem.

Image: Tesla

But the idea that all Teslas built in the past four years have the hardware necessary to achieve Level 5 (L5) automated driving capability was and is a lie, and for some very mundane reasons.

Just a week after Musk made the above pronouncement, I took to a stage in San Francisco, where I was chairing a conference on automated vehicles. George Hotz of had been scheduled to deliver a keynote that morning. I was checking my notifications as I walked into the event venue and saw the headline that Hotz’s Comma One project had been canceled after the company opted not to respond to an inquiry from the National Highway Traffic Safety Administration asking for more information about it.

Needless to say, Hotz did not appear, so I quickly modified a presentation deck on the state of automated driving development and took his place. In the course of my talk, I asked the audience of about 150 people how many thought that Musk could make good on his L5 promise. I could count the number of people who responded in the affirmative on the fingers of one hand, and then proceeded to give my own explanation of why Musk’s claim was a lie.

I think Musk is a true believer in most of the ideas he has brought forth, regardless of how outlandish many seem.

I often get calls from the media for comment on all aspects of mobility technology, including automated driving and electrification, and I’ve repeated this explanation countless times over the years. The fundamental problem is that most people view Musk as some sort of visionary genius and assume he knows what he’s talking about. Thus, when Musk says things that are fundamentally wrong about automated driving, it becomes gospel for people who are coming from a nontechnical background. It has become my job to correct the misinformation.

Before I explain why no current Tesla vehicle is capable of L5 automation, let me give a brief explanation of what these levels mean. The Society of Automotive Engineers published the J3016 standard in 2014, which defined a taxonomy for levels of automated driving capabilities. Level 0 is for vehicles with no driver-assist capabilities at all. The vast majority of vehicles built today fall within L1, with a variety of discrete functions to assist in the driving task, such as lane departure warning or adaptive cruise control.

Level 2 systems combine multiple functions within a control strategy, such as automatic lane centering and speed control, but still require the driver to pay attention and be prepared to take full control at any time. These systems may or may not require the driver to keep hands on the wheel, but their eyes must remain on the road. As of 2020, Tesla AutoPilot is still an L2 system.

Level 5 (the highest level) refers to fully automated vehicles. These are vehicles that are capable of operating in all conditions without any human supervision or intervention. An L5 vehicle can drive itself in any weather and on any roads that a human can. The next step down (Level 4) can do the same, but within a defined operating domain that can be based on geography, weather, or any other conditions. The vehicles being tested by virtually all other AV companies, including Waymo, Argo, Cruise, Voyage, and others, are all L4.

L4 and L5 vehicles can have human controls, and a driver can take control if they desire, but it’s not required. Both of these systems have levels of redundancy that enable them to get to a safe stop if a problem is detected, even when no one is aboard.

The key distinction between L4 and L5 is the ability to drive anywhere and at any time. To do that, the driving system needs to be able to “see” the world around the vehicle, which means the sensors must remain clean and unobstructed at all times.

Ford sensor cleaning. Video credit: Ford

Most people who are critical of Tesla’s approach to AV technology focus on Musk’s insistence on not using lidar. Tesla vehicles are equipped with eight low-cost cameras, along with a single forward-looking radar sensor and 12 short-range ultrasonic sensors. While this solution is unlikely to yield a sufficiently robust automated driving system anytime in the foreseeable future, it’s not impossible, despite the limitations of cameras and the limited redundancy. Even if Tesla can develop software that can reliably understand the world around the vehicle for a safe L5 system, the cars as they are built today will never be L5.

As of 2020, Tesla AutoPilot is still an L2 system

It all comes back to cleanliness.

Those of us who live in regions that have multiple seasons are well acquainted with the challenges of trying to see out of windows caked in slush or getting illumination from headlamps covered in road salt. Even when the weather is warm, billions of insects are killed every year as a result of windshield impacts, and pollen from trees and plants can build up on all manner of surfaces in addition to irritating sinuses.

Waymo sensor cleaning system. Source: Waymo

Ford, GM, and Waymo all have systems on their automated vehicles designed to minimize the collection of anything that can obscure sensors. While washer nozzles, air curtains, and wipers may seem incredibly mundane compared to neural networks and machine learning, it’s often those mundane details that can trip up an engineering project. Remember when a Mars probe crashed because engineers failed to convert some crucial numbers from English to metric units?

If you can’t keep the sensors of an AV clean, it can quickly be blinded. If the virtual driver can’t see, it can’t plot a course through its environment. If that happens, the vehicle can’t drive and is not L5 automated.

A radar sensor caked in slush and winter road grime after 10 minutes of driving. Photo: Author

Of the eight cameras on current Tesla vehicles, three forward-looking imagers are mounted above the rearview mirror and sit within the swept area of the windshield wipers, which could theoretically keep them unobscured. The other five on the sides and rear have no cleaning system at all. As someone who has been driving in winter conditions for nearly four decades, I can guarantee that all five will get obscured with an array of road grime.

However, even those forward cameras aren’t completely safe. When snow and ice build up, the outer tips of wiper blades frequently get lifted up and don’t wipe the complete area, which just happens to be where those cameras are located. Those insects I mentioned? When a mayfly goes splat on a windshield at 70 mph, I’ve yet to encounter a wiper anywhere that can remove it. Don’t get me started on the streaky mess when you try to wipe away bird droppings.

Even in warmer weather, the insects collected on the front of a vehicle can make sensors unusable. Photo: Author

As a self-proclaimed visionary, Elon Musk may not spend much time considering mundane tasks like scraping bugs and dung off windshields, but they are all a necessary part of life. Machine vision is not exempt from the need for a clear view. In fact, it’s more sensitive to such obscurants than the human brain, which can actually do a shockingly good job of seeing around such annoyances.

Until Elon Musk can handle the mundane, Teslas will never have Level 5 automated driving capability.

Sam is a principal analyst leading Guidehouse Insights’ e-Mobility Research Service covering automated driving, electrification and mobility services

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store