The automotive industry is going through the most transformative journey in its history. Only 15 years ago, autonomous cars were a pipe dream; electric cars were the butt of jokes; and safety was reactive, aiming to minimize the consequences of collisions rather than prevent them. Now, in 2024, autonomous mobility services are available to the public across the U.S. and China; electric cars are a desirable mainstream option; and the future of safety is preventative, trying to ensure crashes never happen. This article considers three critical aspects of future vehicles: sensors, software and safety.

Sensors

Radar, cameras and LiDAR form the core trio of sensors synonymous with autonomous driving. All three have been evolving in the automotive industry and continue to progress. LiDAR is coming down to a price point more palatable for OEMs, radar’s performance and utility continues to evolve, and high-resolution camera technologies pioneered for smartphone cameras are being leveraged in autonomous driving systems.

Radar is perhaps the most interesting of the three categories. Automotive radar has come a long way since it was introduced to vehicles around the turn of the century. Back then, radar was good for detecting the distance to the vehicle ahead, and that was about it. Its major advantage was that it worked under almost any weather conditions the car might encounter. Snow, rain, fog, sandstorm—the radar did not care. For fully autonomous driving, however, its resolution simply was not sufficient.

In recent years, advances in semiconductor technology and the emerging demands of the autonomous driving industry have forced radar to evolve. It has sprouted more transmitter and receiver channels (Tx × Rx), growing from 1 × 2 to 3 × 4, and then to 12 × 16 from leading Tier 1 companies and 48 × 48 arrays from leading startups. Large antenna arrays with many transmitting and receiving channels have boosted the resolution of modern radar.

However, transmitting and receiving channels are like pixels on a camera: Simply adding more pixels does not necessarily make a better camera. As far back as 2013, the Nokia Lumina 1020 had 41-MP cameras, yet the iPhone 14 from 2022 uses 12-MP sensors. Which would take a better photo? There is more at play than just the pixel count. Sensitivity, dynamic range and post-processing software make a significant difference to the end image. Likewise, radar performance is not just a measure of transmitters and receivers.

An emerging priority for radar is dynamic range, or the ability of radar to detect both high-reflectivity objects and low-reflectivity objects simultaneously. Consider, for example, a small child standing near a car. The reflection from the car is orders of magnitude brighter than that from the child. For the radar, detecting the child is like trying photograph a grain of rice in front of the sun. It’s very easy for the child’s reflection to be lost to noise.

Advances in semiconductor technology and radar design are improving radar’s dynamic range. For example, at CES 2024, Mobileye showed an example of a radar that could detect a wooden pallet next to a metal railing at nearly 240 meters.

Software

Horsepower is becoming democratized. Scaling an electric motor to double or triple the horsepower is much cheaper than doubling the output of an internal-combustion engine. Cars that would put the supercars of yesteryear to shame are now entering the market at less than US$50,000. In combination with this, electric cars do not possess the differentiating characteristics of engines. They do not have turbo lag or differing exhaust notes; most do not even have more than one gear. In short, future cars will be differentiable not by their performance but by their software.

Software-defined vehicles, connectivity and autonomy also provide a route to perfect safety. The problem with sensors, even the best and most advanced sensors, is that they cannot see around corners or through buildings. No matter how many radar, LiDAR or camera units are fitted to a car, they cannot see that a speeding car on an adjacent city street is about to jump a red light. But connected vehicles could.

Connectivity and software offer a pathway to collective perception, in which cars and infrastructure would share their environmental knowledge with each other. Armed with such information, a connected car could react appropriately to the location and movement of every pedestrian, cyclist and vehicle on the road around it, including those beyond its line of sight. Combine collective perception with a responsible autonomous driver, and a country full of such cars could end road traffic accidents altogether.

This scenario is, of course, a long way off; autonomy on the roads is in the early days, and collective perception is a university research project. However, autonomous vehicles are already starting to show their potential safety benefits.

Safety

In October 2023, the California DMV revoked Cruise’s driverless testing permit, preventing it from operating its robotaxi program and testing cars on public roads in California. The stated reason for this was that Cruise “misrepresented” the safety of the technology. It is clear from this incident and others that AV safety claims will always and should always be subject to public scrutiny. This is not to disparage AVs or their safety, however. Rather, public oversight should be undertaken not with the motivation to eliminate autonomous cars but with the intent to improve them.

Unlike accidents involving human drivers, the intelligence gathered on a single AV collision can generate improvements that can be rolled out via over-the-air updates to the entire fleet. In this case, a collision provides an opportunity to prevent that type of collision from ever happening again, something far beyond the capabilities of human drivers.

By all metrics, the safety of AVs is improving. The data from California shows as much. Going back to 2015, when records were first kept, autonomous cars would barely make it 1,000 miles without the safety driver behind the wheel intervening. Fast forward to 2023, and Cruise submitted 576,000 miles worth of driving without a single disengagement. By comparison, IDTechEx estimates that human drivers in the U.S. are involved in collisions approximately once per 200,000 miles, and in cities like San Francisco, this rate nearly doubles. So perhaps the argument can be made that autonomous cars are already safer than human drivers.

On the other hand, a look through the data on driverless autonomous cars reveals that, on average, a driverless Waymo is involved in collisions once every 52,000 miles. Cruise fares slightly better, with one collision per 63,000 miles. So, they are colliding at roughly twice the rate of average San Franciscan drivers. But again, this is not the whole story. Out of 33 recorded collisions involving a driverless Cruise vehicle, blame can be attributed to the autonomous driving system in only six cases, which equates to once per 344,000 miles. Likewise, Waymo’s autonomous driving system could be held responsible for roughly five collisions, equating to one per 238,000 miles. So by this metric, autonomous drivers are once again safer than the average human driver.

Another metric that should be considered is the rate at which fatalities and serious injuries occur with AVs. Once again, this is lower for AVs than for human drivers.

However, there is a big issue with how this discussion is framed, and it concerns the “average” human driver. Clearly, human driving proficiency is a spectrum, encompassing those who are incapable of staying below the speed limit and off their phones, the super-vigilant types who have never had an accident and everyone in between. The question, then, is with which human drivers an autonomous car should be compared. Are autonomous drivers safe enough when they are safer than 90% of human drivers? Should they be better than every human driver? But how would that be defined? A 17-year-old with a fresh license may never have crashed, i.e., one collision per infinity miles.

Complicating matters is that the true safety record for human driving is not fully known. Not all collisions are reported, and not all dangerous situations or near-misses end in collisions. The industry just about has a grasp of how safe the average person is by measuring the rate at which people crash, cause injuries and cause fatalities. However, insufficient data exists to understand human driving safety across the full spectrum of abilities to the same depth as autonomous driving safety.

It was important to show with data and numbers that AVs could exceed the driving standards of the average human driver. However, to go beyond this and show that autonomous cars are safer than any human driver, the benchmark will need to change or evolve. Even beyond near-misses, there are many metrics that could be used to better understand human driving safety, such as how often drivers go through red lights, how often they tailgate, how often they disobey traffic rules, how often they exceed speed limits and many more. With these metrics, the industry could better understand human driving performance and make solid comparisons to build a case that AVs are safer.

Between sensors, software and safety, it is an exciting time for the automotive industry. The advent of AVs in recent years is an era-defining moment. In 100 years’ time, examples of today’s autonomous cars and robotaxis will be exhibited in museums as part of the story of how the automotive industry followed a pathway to infallible road safety.