#I Took a Ride in a Self Driving Car (and Didn’t Die) – Review Geek

You may have heard of Waymo, Google’s self-driving company, and you almost certainly know about Tesla’s efforts in the industry. But I didn’t take a ride in either of those cars. Instead, I rode in a heavily modified car with hardware designed by Ambarella. This company specializes in imaging processors that can identify people, cars, and more.

While plenty of devices—like smart cameras—promise that capability, most of them rely on the cloud to accomplish the task. Ambarella doesn’t focus on the cloud. Instead, it specializes in an “edge” application that executes all the identification locally. By performing the task locally, Ambarella can dramatically speed up results, which in a moving vehicle is incredibly important. The company started in dashcams, and you’ll find its processors in Nextbase cameras and more.

In the past, getting all that processing done required a host of processors, which increased resource requirements like room for the hardware, cooling, and energy usage. At CES 2022, Ambarella unveiled a new CV3 AI SoC (system on chip) that can accomplish more with less hardware. The company complains a single chip can manage level 4 self-driving tasks where multiple chips would have been required in the past.

Alas, it’s too soon to test with that hardware, so the self-driving car I rode in used the previous generation “multi-chip” system. On top of the vehicle and the sides, I could see quite a few cameras and radar systems, and the inside housed several monitors.

Some of those monitors were for my benefit; they gave me a view of what the car “saw” and how it identified people, objects, and the road itself. But one additional monitor looked like, and served as, a digital “rearview mirror.” The view it gave seemed more valuable than the traditional review mirror next to it. I could see more going on behind us in the digital version.

So how was the ride? Well, somewhat short for necessary reasons. I was at CES, had other appointments to get to, and no doubt Ambarella had other journalists waiting to take a ride. We took a drive more or less around the block. We stopped at a midway location, parked, then came back. It was around a 15-minute ride.

At first, the monitors were a bit tough to understand. But that’s because the car generated a map of its surroundings in real-time, inside the parking lot area. It quickly got a layout, and I saw marked parking spaces, parked vehicles, pedestrians, and more. But the markings jumped around a lot; it wasn’t a stable view. Still, it was interesting to see the car essentially map its area as it drove.

When we reached the intersection to pull onto the road, the car spotted the merge point, noticed the stoplight, and obeyed. Eventually, it chose to turn onto the road, and that’s when the view on the monitors shifted. Ambarella had full maps pre-loaded for the road, making the car’s job easier.

As we drove along, the car quickly and accurately marked lane dividers, where the sidewalk met the road, pedestrians, other vehicles, and more. It noted intersections and stopping points and determined what to do. The car stopped on red and kept going through green, but it was the edge cases that were impressive to watch. On multiple occasions, the vehicle stopped at an intersection and then made a new determination. In one instance, it correctly identified a “no turn on red” indicator and waited until the green light. But when it was legal, it turned right on red when there was room.

When we reached the two destination parking lots, it identified an empty slot and parked itself by backing into the space. The Ambarella crew let me know that the vehicle A.I. essentially chooses a spot at random, without any real rhyme or reason. That seemed fine in the setting we tested as all the parking spaces were reasonably close to an entrance. But that might not be so great at, say, a grocery store, where the car might choose a location far from the building.

So we managed to drive to another parking lot, park, then return and park again. And the car even navigated stop lights, street merges, and entering or leaving the main road for the parking lot. But you might be wondering how well it, well, drove.

The closest analogy I can think of is it had the skill level of a teenager learning to drive. Which I suppose is impressive but didn’t always feel quite that way. Every turn was a bit too tight, every acceleration or brake was too fast. It honestly felt like an unsure teenager still trying to master safe lane changes and braking distances. With one exception—the parking job. To park, the car chose to execute a three-point turn and back into the space. And by golly, it did that smoothly even when a gaggle of people crowded the spot next to the car’s destination.

In theory, as Ambarella lets the car drive itself more, it will get better and better at what it does. Other outfits like Waymo may give a smoother ride, but I haven’t ridden in one to be sure. And in fairness, Amberella’s primary focus is the hardware behind self-driving, and everything I saw suggests it’s more than capable of getting the job done.

But I would be remiss if I didn’t mention the two occasions when our safety driver intervened. On the first occasion, I couldn’t say for sure the car “failed” and intervention was required. We were driving along the road, and another vehicle pulled out of a parking lot directly in front of us with absolutely no room to spare. That driver made a serious judgment error, and we definitely could have hit them.

Our safety driver immediately slammed the brakes as soon as we saw the car. I would have as well; instinct just takes over. I’m not sure the self-driving car had the opportunity to brake itself; for all I know, it tried to, but manual control took over. I think the driver made the right choice, and it’s certainly better to have a journalist write “safety-driver took over just in case” than “self-driving car failed to stop in time and we hit an idiot driver.”

But the second time the safety driver hit the brakes was scarier. On this occasion, we were at an intersection getting ready to turn right. We had a stoplight, but turn on red was allowed. The car crept forward, rather human-like, in anticipation of finding an opening in traffic. But there was no such opening, cars kept coming, and the self-driving car kept creeping closer and closer to oncoming traffic. At one point, I felt the engine rumble as it planned to accelerate, and the only “hole” in the traffic did not look big enough. The safety driver hit the brakes to stop the car from merging.

An Ambarella rep who rode in the car assured me that he didn’t think we were going to attempt the merge and that the braking was probably unnecessary but technically required by safety policies. I’m not convinced. It felt like the car started to merge, and as it did, the driver braked (and, again, I would have, too).

It’s worth repeating that Amberella’s primary goal isn’t to develop the software that drives your car. It creates the hardware that turns camera images into useable data for the software to use. It’s entirely possible the very companies Amberella will sell its chips to would have fared better. And at the end of the day, I took a ride in a car that drove itself successfully for nearly the entire trip.

Obviously, “nearly” isn’t good enough to start putting in every car on the market. And the hardware costs for cameras and radars alone puts those kinds of features at the upper tier of vehicle purchases, But I still walked away impressed. Heck, I walked away!

If you had asked me when we might see proper level four self-driving in cars for sale a few weeks ago, I probably would have answered ten or twenty years. Now I’d call ten years the upper level of that guess. Possibly quite sooner. I can now envision a future where my son never learns to drive: because the car did the learning instead.