The Battle for Best Semi-Autonomous System: Tesla Autopilot Vs. GM SuperCruise, Head-to-Head

Being the best isn't just about technological prowess, it's about how well the designers understand human driving—and human nature.

Alex Roy

The stench of clickbait around self-driving is so thick these days you need a windmill and smelling salts to survive the news cycle. I very much want to believe a traditional automaker can innovate, but I've got bad news: self-driving theater is the new security theater.

When was the last time a traditional carmaker really innovated? I'm not talking about fuel economy gains that evaporate under government scrutiny and threats of jail time; I'm talking about demonstrably superior technology unique to one company that is months or even years ahead of the competition.

Whatever one's opinion of Tesla Autopilot, there is no question that Tesla won the first battle in the war for self-driving. Though still no more than a semi-autonomous driving system, Autopilot was the first commercially available system to demonstrate what was possible, good and bad, for semi-autonomous driving systems. The entire auto sector has been playing catch-up since Autopilot's release in October of 2015.

Until now, that is, because GM SuperCruise has arrived, and it's both innovative and brilliant.

Cadillac Super Cruise
Cadillac

Is it the perfect semi-autonomous driving system? No, but SuperCruise is an epic technological achievement in the major ways that other pretenders weren't. Remember Mercedes-Benz Drive Pilot? It was last summer's alleged Tesla-Killer. It sucked so badly I had to write not one but two deep dives to explain why. It arrived with a marketing campaign—since buriedso misguided it made Tesla's Autopilot nomenclature seem like the height of clarity.

What about Volvo's Pilot Assist I/II? Better than Drive Pilot, but nowhere near as good as Autopilot. Why didn't Volvo rouse my ire the way Mercedes-Benz did? Because Volvo didn't launch a marketing campaign calling their vehicles "self-driving," like Benz did.

I'm not going to name the others. They should bury their respective systems until they have something as good as Autopilot—or, now, SuperCruise.

What Are We Actually Comparing?

GM SuperCruise is currently only available in the (underrated) Cadillac CT6. Tesla Autopilot is available in the Model S, Model X, and Model 3. This comparison is between SuperCruise and the Model S / Model X iteration of Autopilot, despite the fact that I just completed a 2,860-mile test drive of a Model 3. This is because 1.) I believe the Model 3 user interface is sufficiently different from the S/X to constitute a different system, despite the same software; and 2.) in my opinion, the iteration in the Model S/X has the superior user interface. As a result, I'm not comparing SuperCruise to the newest iteration of Autopilot, but to what I consider the best version.

Note: Technology is only as good as our understanding of it, and that understanding starts with examining one's expectations of the system in question. Technology can solve a lot of problems, but it can't solve for human nature; give people any tool, and some will use it in unforeseen ways. Therefore, a major criteria for a safe semi-autonomous driving systems is for that system to not exacerbate bad driver judgment. Which means that the answers to questions like, "Which system is best?" and "Which system is safer?" depends somewhat on the driver.

Defining Semi-Autonomy

The problem with comparing semi-autonomous driving systems is that there is no consensus on what they're supposed to do. The Department of Transportation's meager contribution to the debate was to adopt the Society of Automotive Engineers (SAE) definitions of automation, which are so vague that no automaker wants to definitively say where its cars fall on the continuum. Automakers can't even agree on the necessary hardware. And if hardware were everything, the newest E-Class, with its rear-facing radar, would have beaten the Tesla Model S in my comparo from last year. 

Plus, the SAE levels have virtually nothing to say how about how hardware and software should work together. For now, we're stuck with multiple systems at or just below SAE Level 2, called "Partial Automation" and defined thusly: "The driving mode-specific execution by one or more driver assistance systems of both steering and acceleration/deceleration using information about the driving environment and with the expectation that the human driver performs all remaining aspects of the dynamic driving task."

Translation: The car can drive itself, but the human is fully responsible for anything that happens, which includes taking over when the system disengages. Because it will.

Let's break down the differences between the two best systems on the market: Tesla Autopilot versus GM SuperCruise.

Alex Roy

SuperCruise in its operational domain.

Operational Domain

This is the technical term for when and where the semi-autonomous system will function. Tesla Autopilot will theoretically work anywhere, provided the forward-facing camera and radar can "see" enough. That bar is flexible and fairly low, determined by clarity of lane markings, weather, and (apparently) proximity of other cars. Tesla is vague about this, and often errs on the side of What the hell, let's give this a shot. This is both Autopilot's strength and weakness. Within a few hours of first use I had a good sense of where Autopilot would engage—and, importantly, stay engaged—which is key to safe and pleasurable use. 

I loved it, and still do, but there is an obvious issue. Many of the high-profile accidents that occurred while Tesla Autopilot was engaged appear to have been the result of overconfidence in the system; i.e. the related assumptions that: 1.) if engaged, Autopilot will stay engaged; and 2.) when engaged, Autopilot can safely navigate most if not all hazards that would occur within its operational domain. If a system's operational domain is theoretically global, and if people don't understand that system's limitations, bad things will happen.

Cadillac is the first company to deploy a real solution for this industry-wide problem.

GM got a lot of grief for SuperCruise's "late" arrival to the semi-autonomous driving game, but the wait was worth it, because the company first took time to deploy a fleet of Lidar-equipped vehicles to map not only the U.S. Interstate system, but also portions of Canada and China. Lidar is the most accurate way to gather hi-resolution road data, and SuperCruise's operational domain is wisely limited by GPS to what they call "limited-access" roads, otherwise known as divided highways, which use on- and off-ramps instead of intersections. This restriction, and Cadillac's communications strategy around it, is clearly designed to mitigate the likelihood of an accident like Josh Brown's, which was the result of Brown engaging Tesla Autopilot on a two-way highway, then failing to intercede before his Model S struck a tractor-trailer that was making a left turn across his path of travel, and which Autopilot failed to recognize.

GM's mapping solution isn't perfect, though. In and around Los Angeles, SuperCruise was excellent, working exactly where I expected it to, and disengaging as I approached exits and interchanges—most of the time. I quickly learned that if you're on the I-10 eastbound and want to exit left onto the northbound 110, SuperCruise inexplicably remains active seconds after I expected it to have disengaged; e.g. clearly after the transition from highway driving to beginning to exit the highway. (SuperCruise will not exit a highway of its own accord.) Seconds of uncertainty seem like an unsafe eternity when using these systems. 

In and around NYC, it was almost impossible to engage SuperCruise, even under excellent conditions. It worked eastbound on the Long Island Expressway but not westbound, and it barely worked on the Grand Central Parkway. Then, inexplicably, the little gray steering wheel icon on the dash lit up on the FDR Drive northbound, indicating SuperCruise was available. I activated the system; it worked for about six seconds before disengaging. This happened twice more. GM has some work to do.

About 18 months ago, I took a Tesla with first generation Autopilot from Battery Park to 178th street—the full length of the FDR Drive—with only two disengagements. Maybe SuperCruise has a higher safety threshold than Autopilot. I don't know. I do know that the FDR Drive appears to fall within SuperCruise's operational domain; that the day on which I was driving was cold and clear (meaning no obvious system visibility issues due to weather); and that the display showed both lane markings set to green, meaning the SuperCruise recognized them. SuperCruise's operational domain is supposed to be clear, but if it won't engage when expected, or if it disengages when I don't think it should, I want to know why.

Alex Roy

SuperCruise's brief engagement on the FDR Drive.

Given how Cadillac is marketing the system, they need to make their SuperCruise-approved road map public. Better yet, they need to make such a map the default overlay for their in-dash navigation system, with supplemental icons in both the main cluster and heads-up display. Tell me when and where my technology works, before I buy, and I'm your customer. Exceed my expectations, and I'm yours forever. If you disappoint me, or don't tell me what I need to know in advance, we've got a problem.

Winner: GM SuperCruise, by a hair

Updates

In theory, Tesla is only one acquisition or strategic partner away from being able to define Autopilot's domain in a serious way. They've already announced a big impending map update, but there's no evidence it's domain-related. That's the thing about Tesla and updates: their frequent over-the-air (OTA) software updates mean that, if there's a problem Tesla can solve with a download, they will.

Cadillac? Not so much.

Alex Roy

Cadillac CT6 Map Display

GM claims OTA map updates will occur quarterly. Why only quarterly? I'm guessing they have [X] Lidar-mapping cars, and [Y] amount of time. X/Y = quarterly updates. That's pretty good in theory, but not so good if a road closure occurs the day after your SuperCruise map update. Because SuperCruise is so consistently good in its domain, one's confidence in it is far higher than in a Tesla at any given time. If SuperCruise thinks a road falls within its operational domain but conditions on the ground have changed and the road no longer qualifies, once-reasonable expectations for the system are rendered unsafe.

Example: Here's what happens when unreasonable expectations for meet an unanticipated construction zone.

Speaking of OTA, Tesla also gathers driver behavioral data from owner vehicles via what they call Fleet Learning, which is used to improve Autopilot over time. GM does nothing of the sort. Their vastly larger pool of customer cars aren't designed to gather or send data back to the mothership, which limits SuperCruise development to simulation plus whatever data is gathered from company-owned cars. How many company-owned cars has GM deployed to gather data? A dozen? A hundred—maybe a thousand? Most likely it's far fewer than the nearly 80,000 Fleet Learning-capable vehicles Tesla delivered to customers in 2016 alone.

Alas, SuperCruise behavioral updates remain hobbled by those pesky dealer franchise agreements. If you want any aspect of SuperCruise updated other than your maps, you need to take your car to a dealer. It's the dealers that are throttling OTA more than any other single entity.

That GM is even considering updating SuperCruise's software over time is a very big deal, because there's no indication Mercedes is updating Drive Pilot, even at the dealer level. Is anyone else doing OTA? Nobody's talking. Sad!

Good on GM. But I still don't want to go to a dealer. Dealers are yesterday. OTA is tomorrow, today.

Winner: Tesla Autopilot

Mode Confusion

This is the view out over the windshield of a Tesla Model S on Autopilot:

TechCrunch

How do you know Autopilot is engaged? The tiny steering wheel icon to the right of the speedometer turns blue. That's it. I've been a huge fan of Autopilot since Day 1, but this remains one of its biggest weaknesses: it's way too easy, especially on long trips, to forget whether or not it's engaged.

This is what SuperCruise looks like once engaged:

Alex Roy

This is what it looks like if the driver chooses to manually change lanes:

Alex Roy

This is what the initial warning looks like when the system is about to disengage:

Alex Roy

This is what it looks like when disengaged:

Alex Roy

"Perfection," according to Antoine de Saint-Exupery, "is not achieved when there is nothing more to add, but when there is nothing left to take away."

I cannot imagine a more elegant or obvious solution to possible mode confusion than Cadillac's. 

Winner: GM SuperCruise

Engaging the System

Tesla Autopilot is easily engaged in a Model S/X:

Alex Roy

Once the little grey steering wheel icon appears to the right of the speedometer, pull twice on the cruise control stalk left of the steering column. Two ascending tones chime, and voila! Autopilot is engaged.

If only Cadillac's UI was so simple:

Alex Roy

In order to activate SuperCruise, radar cruise control—which is controlled by a typically loathsome GM kludgecluster of buttons on the left of the steering wheel—must already be engaged. Once the little grey steering wheel icon appears to the right of the speedometer, press the SuperCruise button in the center of the cluster, and voila! SuperCruise.

Well, sometimes. It's also possible that you press the wrong button, thereby disengaging cruise control, then have to reengage cruise control with a different button, then press the SuperCruise button. I can't tell you how many times this happened.

Winner: Tesla Autopilot

Disengagements / Transition Warning System

Disengagements can be good or bad. The good ones are those you choose, because you want to take over control of the vehicle. These are easy: tap the brakes or move the steering wheel more than a few degrees, and you're back in full command. You made the choice to take command of the car, so there's no ambiguity. 

Then there are the type of disengagements that you don't choose. The method by which a semi-autonomous system warns a user of a forced disengagement is called a transition warning system (TWS); since either system can disengage literally at any time, it's critical that transition warnings are as obvious as possible.

Both systems use a combination of increasingly intense visual and audible warnings to alert the driver to an impending disengagement. The Tesla's dashboard flashes white, then red. The Cadillac's engagement bar flashes red. Neither system's audible warnings are loud enough for my taste, but they are louder than most of the other systems I've tested, which are downright dangerously quiet. A Tesla vehicle has the advantage over a Cadillac CT6 in that it's electric, and therefore has a lower noise floor; the SuperCruise warnings seem louder, but are somewhat masked by road and engine noise. 

Cadillac's one useful advantage over Tesla, though, is an aggressive seat vibration warning system in addition to the audio and visual alerts. When it comes to transition warnings, more and louder are better.

Here's an idea perfect for SuperCruise: Since GM knows SuperCruise's operational domain, why doesn't the system give the driver a warning a minute or so before the vehicle crosses the disengagement geofence? How about including some verbal warnings? 

Winner: GM SuperCruise

Hands-On, Hands-Off

A totally unfair comparison. Tesla's current generation Autopilot isn't like the old one: it now requires your hands on the wheel every few minutes. If you ignore the warnings and trigger three disengagements, the system will remain disengaged until you pull over and come to a full stop to reset. GM SuperCruise? A true hands-off system by design. Enter its operational domain and engage it, and you can just sit there monitoring the situation.

Tesla's first-gen Autopilot was a fairly good hands-off system, despite the company's denials it was intended as such. In perfect conditions, I observed the system remain engaged, hands-off, for as long as 30 minutes. Stick some water bottles into the steering wheel and it was possible to stay hands-off much longer.

Tesla Autopilot
Tesla

But those days are over. Tesla would argue that their system isn't intended to be hands-off. I've even defended their use of the name "Autopilot." But we can't ignore how good the system feels when it works, and how compelling it is to take one's hands off the wheel, even briefly. That the first generation was interpreted as hands-off, and the new one is officially hands-on, suggests that Autopilot's hands-off safety is limited by software that isn't here yet, and a piece of hardware Tesla doesn't currently offer.

A hands-on system may sound safer on paper, but in reality a hands-off system negates a kind of driver-vehicle mode confusion. With a hands-on semi-autonomous system, the car will still make driving decisions—say, steering you to the center of your lane via lane-keep assist—even when the driver is engaged with the car via the steering wheel; this muddies the issue of whether the car or the driver is ultimately in charge. With a hands-off system like SuperCruise, there's no confusion: if your hands are on the wheel, you are driving the car, and the myriad safety systems are backing you up; if your hands are off the wheel and SuperCruise is engaged (which, again, is always visually obvious), the car is driving itself and you have become the safety back-up. Simple and elegant.

Winner: GM SuperCruise

Driver Monitoring System

This is GM's triumph, and Tesla's major weakness. Given the low and declining state of driver education in this and other countries, it simply isn't possible to guarantee the safety of a semi-autonomous system without a driver monitoring system (DMS). If it were, Tesla wouldn't have had to shorten Autopilot's hands-on time interval, or add the three-strikes rule, for the second generation system. 

To be fair, you can't guarantee safety even with a DMS. But it sure helps in a big way. In fact, camera-based driver-monitoring systems should be standard on all cars, even those lacking semi-autonomous features.

This is what Cadillac's brilliant DMS looks like:

Alex Roy

The driver's face is illuminated by six infrared lamps positioned on the steering wheel at 11 and 1 o'clock, while an infrared camera on top of the steering column monitors the driver's face. Turn your head or take your eyes off the road for too long—which is not that long at all—and disengagement warnings will begin.

I tried to trick it. I held my phone and began texting, first in front of the steering wheel, then off to the side. I wore four different pairs of sunglasses. No matter what I did, the disengagement warning went off. I even tried something innocent and reached for something I'd dropped on the floor of the passenger footwell. No go. The system knew. The red bar on the steering wheel started to flash, and I grabbed the wheel to keep the system engaged—though I didn't actually have to go that far. If the driver simply refocuses his attention to where it's supposed to be, the warnings stop and the system remains engaged, with no need to make contact with the steering wheel.

And, to be clear, I don't want to be able to trick the system—it should absolutely know if I'm not holding up my end of the bargain. That's why it's there.

Meanwhile, with Autopilot, if you keep your hand gently on the wheel you can pretty much text away. Autopilot has been the state-of-the-art since its release a little over two years ago, but its driver-monitoring system is one of the few areas where Tesla isn't light years ahead of the competition.

Winner: GM SuperCruise, by a mile

Lane-Changing

Tesla Autopilot will change lanes for you. SuperCruise won't. Should be an easy win for Tesla, right? Wrong. In the absence of rear-facing radar, Tesla uses short-range ultrasonic sensors to determine whether there are any cars in close proximity. If you don't feel like looking over your shoulder and trust that there isn't a car in the neighboring lane closing at high speed, you can use the turn signal indicator to request an automated lane change. With some luck, you will survive this decision 99 percent of the time.

Tesla's system works perfectly—assuming you look over your shoulder first. Let's be serious: most people don't know what an ultrasonic sensor is, let alone know how far it can see. Yes, the manual and the law both say the driver remains responsible; no, that does nothing to clue anyone into the fact that ultrasonics have insufficient range to safely allow for blind lane changes at interstate speeds. 

Furthermore, if one makes a manual lane change with Autopilot engaged, it immediately disengages the system, forcing you to reengage it after the lane change. This is annoying.

SuperCruise eliminates both of these problems. Drivers must always make the lane change manually, but instead of disengaging, the system goes on temporary stand-by mode. All you have to do is place your hands on the CT6's wheel and move it slightly, and the engagement bar turns blue, meaning you have control of the car. Change lanes, center the vehicle, pause: the engagement bar returns to green. Release the wheel. 

It's wonderfully convenient. Plus, you're far more likely to look over your shoulder and check your mirror if you're the one making the lane change. Until a truly safe automatic lane change system arrives, this is as perfect as it gets.

Winner: GM SuperCruise

Situational Awareness

Given that both Tesla Autopilot and GM SuperCruise can disengage anytime, and neither transition warning system is truly adequate, situational awareness is everything. It would therefore make sense for automakers to share with drivers as much information as possible about what the car actually sees.

Tesla was the first company to take situational awareness to another level. The following images are from a Model S with first-generation Autopilot hardware, delivered approximately 12 months ago.

Alex Roy

Model S with first generation Autopilot hardware.

The above shows that the camera can see the left-lane marking; the car directly ahead; and multiple cars around it. The Autopilot steering wheel icon appears greyed out, indicating that the system is ready to be engaged. From experience and based on what the car can see, I would have confidence that the system will remain engaged.

Here's my situational awareness once Autopilot is engaged:

Alex Roy

Model S with first generation Autopilot hardware.

I'd call this a medium confidence scenario. The left lane marker and car ahead have turned blue to indicate Autopilot sees them. If the right side lane marker were blue, I'd be at the highest state of confidence.

Here's what a low confidence scenario looks like:

Alex Roy

Model S with first generation Autopilot hardware.

I call this a low-confidence scenario because neither lane marking is illuminated, which suggests Autopilot is relying solely on the truck to remain engaged. If the truck changes lanes or chooses to exit, Autopilot may choose to disengage, or it may choose to follow it. I've experienced both.

I love Tesla's system, but it's inconsistent when identifying two-wheeled vehicles and also requires a level of user interpretation. Still, it's light years ahead of everything else on the market. If only Tesla offered a real explainer to first-time users.

Interestingly, Tesla vehicles sold since October 2016* use different hardware, and the Autopilot interface in these "HW2" cars no longer display vehicles in neighboring lanes—plus, the "locked on" vehicles are now silver and all other vehicles, even trucks and motorcycles, are depicted as cars. Why Tesla changed the volume and nature of Autopilot information displayed in HW2 cars is fodder for another article, but I think the older and more information-rich system was superior.

So, what about SuperCruise?

Alex Roy
Alex Roy

Instead of a nice big central display, like in a Tesla, we get something on the left similar to Mercedes-Benz's inadequate Drive Pilot display. It shows lane markings and a vehicle ahead, but tells us nothing about the system's confidence. The display doesn't show distance, or vehicles next to or in front of it in a meaningful way. It's binary: the system either indicates there's a car ahead, or it indicates nothing. Likewise, the lane markings: it sees two, or it sees none. What happens if I look out over the dash and observe one lane marker partially (or fully) disappear? Does that mean SuperCruise might disengage? I still don't know how SuperCruise makes that decision, because I experienced multiple disengagements within its operational domain and under conditions I deemed at least good, if not excellent.

GM does offer one killer feature Tesla lacks, which is an excellent night-vision camera. Though not part of SuperCruise, it offers a clear image of traffic ahead. It can't see through or around cars, but it clearly displays people and two wheelers in bright bounded boxes:

Alex Roy

In LA, I twice spotted broken-down vehicles with people changing tires close to the roadway. I initiated a manual lane change and then continued within SuperCruise. I wouldn't have seen them without night vision.

Night vision adds another dimension of situational awareness Tesla lacks, but it doesn't work all the time. Although Cadillac states it won't work in daylight, it did for me, part of the time. It's really too bad this isn't integrated into SuperCruise, because there's a lot of good here.

My initial take was to hand SuperCruise the category win, but Tesla's situational display is available 100 percent of the time—and it's useful even when Autopilot isn't engaged, which makes human driving safer, too.

Cadillac's terrific night vision display isn't the default setting in the gauge cluster. For a such a useful safety tool, this is insanity.

Winner: Tesla by a hair when Autopilot is engaged—and by a mile when it isn't.

Radar Cruise Control & Cut-Ins

One of the most frustrating things about radar cruise control systems is how often they allow vehicles to cut into the gap between your car and the one you're following. Even if you adjust the follow distance to the minimum setting, there may still be enough room for a car to cut in. This is a function of your forward-facing radar's field of view, which determines how early it can detect a cut-in, and your cruise control software, which determines when to slow down and by how much.

Now we get into one of the really grey areas of semi-autonomous systems: Is it better to allow the shortest follow setting to be notably close—like Tesla's—to reduce potential cut-ins? Or is it better for the minimal follow distance to be a little longer but with the threat of more frequent cut-ins, as SuperCruise is set up?

Let's add another factor. Teslas are electric, so even their smallest battery pack model is more immediately responsive than an internal-combustion CT6, which means that no matter what the behavior of the car you're following, your Tesla will maintain that minimum distance without the slight delay inherent to the CT6, with its 3.0L twin-turbo V-6. That extra delay increases the likelihood of cut-ins, which leads to another problem: whenever SuperCruise disengaged because of a cut-in, it didn't seem to want to re-engage, even within its operational domain and under excellent conditions. This made using SuperCruise annoying even in light traffic. It moved from a safety feature to an annoyance feature.

SuperCruise has one more minor issue, which is really more of a function of the CT6 as well as pretty much any internal-combustion car with radar cruise control: the transmission. That is, because the behavior of the car you're following determines your speed, if they're erratic you might find yourself in a weird gear-hunting purgatory until the car in front of you settles on a consistent speed. This was annoying in the CT6, though less so than in the Mercedes E-class.

Dear GM: a little transmission software upgrade is in order. Do I have to go to a dealer to get it? Ugh.

Winner: Tesla Autopilot

Lane Keeping

SuperCruise's Lane Keeping Assistance System is my new favorite. Steering inputs are gentle. It feels like it knows the roads, which might make sense based on Cadillac's Lidar mapping efforts. I never felt nervous or unsafe, even when it made mild corrections.

Well done, GM. This is what innovation looks like.

Tesla Autopilot remains excellent, when it works, under ideal conditions, but compared to SuperCruise it feels a tad more nervous, making more course corrections than I would like. Of course, Autopilot gets updated so frequently it may already have improved over the version I drove recently. (Even as I was writing this, Tesla released yet another update.)

Everything else I've ever tested ranged from mediocre to plain unsafe, with one exception: George Hotz's Comma.ai prototype, which even in an early iteration felt at least as good as first generation Autopilot. I only drove it briefly on the I-405 south of San Francisco, but it was very impressive.

Winner: GM SuperCruise, by a hair

Which System Is Best?

If you have any doubts about your attentiveness behind the wheel, neither system will solve it. These offer convenience with safety attached—but that safety depends on you dutifully fulfilling your role as a back-up to the system.

I cannot stress the overwhelming safety advantage of Cadillac's driver monitoring system. That any semi-autonomous system doesn't include this is outrageous. Still, this alone doesn't tell the full story, because a truly safe semi-autonomous driving system would have to combine Cadillac's DMS, maps, and steering wheel display with Tesla's situational awareness, user interface, and radar cruise control behavior.

Which is why, for now, the answer as to which is best depends on you.

For the average person looking for a true hands-free system, SuperCruise is the only way to go. An easy learning curve, defined operational domain, and great DMS mean SuperCruise will be safer to engage, and safer when engaged. Frustratingly, that will be a lot less than Tesla Autopilot.

Tesla owners and those willing to climb Autopilot's learning curve are going to feel constrained by SuperCruise's restrictions. Since Autopilot isn't a true hands-free system, its safety is ultimately determined by the user's real-time understanding of its limitations, which are a lot less clear than with SuperCruise. Master it, and you will love it.

But don't let "self-driving" headlines fool you. People died thinking seat belts would save them. Then it was ABS. Then traction control. Then stability control. Then airbags. And now, this. Until real self-driving cars arrive, you need to know not just how to drive, but how your driver assistance systems drive.

If semi-autonomous convenience is your sole criteria, the final decision isn't about the learning curve, but geography. If I lived in the LA basin, SuperCruise is the clear winner. But I live in NYC, where Autopilot was more effective, if only because SuperCruise often refused to engage in places where Autopilot worked like a charm.

I strongly recommend an extended test-drive of both systems in your own real-world conditions before making a decision.

The Triumph and Tragedy of GM SuperCruise

SuperCruise is everything GM promised, and the first triumph in semi-autonomous driving to come out of a legacy car maker. The tragedy of SuperCruise is directly related to how excellent the system actually is, because in the three months since its release every review has been stellar, and yet while Autopilot warrants its own Wiki, SuperCruise gets no more than a single mention at the bottom of Cadillac's CT6 page.

There is genius inside GM, but apparently you have to go looking for it.

SuperCruise is an idea executed too well by the company that created it, and apparently already abandoned by management whose attention is elsewhere. GM's latest investor deck didn't mention SuperCruise once, suggesting management doesn't understand it—or worse, has already given up on it.

"We are disrupting ourselves," General Motors CEO Mary Barra said in November of 2015, and boy did she keep her word. Perfect being the enemy of good, Barra doesn't seem to realize that disruption cuts both ways. While she and the rest of GM's board don kneepads each morning before begging Kyle Vogt, of the Cruise Automation division, to deliver a fully autonomous car, the kernel of the world's best semi-autonomous driving system sits, already unappreciated, in a single fantastic and underrated Cadillac model.

Cadillac, and GM, should be shouting SuperCruise to the heavens and advertising absolutely everywhere. The fact that they aren't does not bode well, because it suggests the company at large doesn't even realize just how good, or relevant, or marketable SuperCruise actually is.

Mary Barra, Give Innovation a Chance

No one knows when Cruise Automation—or anyone else—will deliver Level 4 self-driving cars. But between now and then, GM is one decision away from the being the top car company in the world, if only they would make full use of the technological jewel already in their hands.

SuperCruise is currently a $5,000 option on the CT6, which is outrageous given the totally commoditized hardware required. If GM wants to earn back their R&D, drop the price to $1,000 and make it optional on each and every GM car currently on the road.

GM needs to act. The rest of the industry is watching. It would be a tragedy if SuperCruise failed because GM didn't market or develop it further. Does anyone really believe Tesla won't install driver monitoring systems on future models? Everyone will. Because they'll have to. GM has proven this.

Or, if GM really cared about safety, they would make SuperCruise free and change the world in the process. But that might be a little too disruptive.

Alex Roy is editor-at-large for The Drive, host of The Autonocast, co-host of /DRIVE on NBC Sports, and author of The Driver. He has set numerous endurance driving records in Europe and the U.S. in the internal combustion, EV, 3-wheeler, and Semi-Autonomous Classes, including the infamous Cannonball Run record. You can follow him on Facebook, Twitter, and Instagram.

*This article has been corrected to reflect when Tesla changed over its hardware. It was October 2016, not January 2017, as originally written.