“Is this the panic button?” I ask, staring at a cartoonishly large red plunger installed on the center console of a Lexus CT 200h.

“Actually yes,” replies Bryan Galusha, cofounder and safety driver from Point One Navigation, a precision location startup that sells souped up GPS rigs to autonomous-vehicle companies. The button disconnects the self-driving system and restores the vehicle to human control.

It’s a cloudy spring day, and we’re in the pit area at Thunderhill West Raceway in Willows, California, where, for the fourth year running, self-driving-car companies are gathering to test out their wares on a closed racecourse. The event is organized by Joshua Schachter, founder of early social-bookmarking service Del.icio.us and also of MemePool, a noted compendium of early internet weirdness. Now he mostly invests in early-stage companies, including many of the outfits at the track today. The event is meant to be a fun meetup of nerdy solidarity, rather than an official proving ground, but the winner will certainly take home bragging rights.

Joshua Schachter invests in early-stage companies and organized the gathering of self-driving-tech companies at hobbyists at the Thunderhill Raceway this spring.

David Robert Elliott

Aaron Nathan, the CEO of Point One, is sitting in the back seat of the Lexus, cradling his laptop. Peering at a column of text on the left of the screen (the rightmost column has what appears to be C++ code), he makes a small change, saves it, and says, “Let’s go!”

Antonio García Martínez (@antoniogm) is a writer and Ideas contributor for WIRED. Previously he worked on Facebook’s early monetization team, where he headed its targeting efforts. His 2016 memoir, Chaos Monkeys, was a New York Times best seller and NPR Best Book of the Year.

Galusha guides us through the first few turns while running down the checklist taped to the steering wheel with Nathan. Before they switch to self-driving mode they make sure every system is on and functioning, almost like a pilot preparing for takeoff. Load route: check! Load mission: check! Verify heading is valid: check!

List done, Galusha takes his hands off the wheel, and it starts moving itself. There is something almost supernatural about this moment that requires a leap of faith, but soon the car is zooming down the track, guided down to 0.1-meter accuracy (hence the company’s name).

In track racing, there’s a concept known as the line: the theoretical perfect path around the course that minimizes total distance, maximizes speed on turns, and involves just brushing the car’s tires against the edge of the inside curve (the so-called apex). An optimal path somewhat resembles a pinball flying through a narrow track and hitting every apex of every inside corner along the way.

Earlier, alongside the Point One pit crew, we had watched on a map as the dot representing their autonomous Lexus described graceful arcs in accelerating laps as the company’s engineers pushed it faster and faster through the twisty course's 10 turns. These were the first runs of the day, meant to shake out any bugs, with a single human inside in case something went wrong.

PointOne isn’t really a self-driving-car company per se; the vehicle was just following hyper-precise waypoints along the track. The company's secret sauce, and the reason its GPS is much better than the one on your phone, is a network of proprietary beacons installed throughout the western US that corrects for the errors implicit in civilian GPS signals, tightening the margin of error to a few finger widths.

Back inside the Lexus, the autono­mous system is still taking us through every turn, but more slowly than strictly necessary. The track safety officer didn’t like the idea of our onboard photographer’s camera turning into a head-bashing projec­tile in the case of a crash, so he’d ordered the team to keep things slow.

Then comes turn 7. An almost 180-degree hairpin, this turn requires going wide to the right of the track, then cutting sharply to the left, hitting the tip of the wedge-like apex, and then proceeding to a sweeping turn 8.

“This one’s going to be a little fast,” Nathan announces from the back.

The car veers widely to the right, setting up the hairpin, but then doesn’t brake as it turns abruptly left and maintains its full straightaway speed.

Elijah Nouvelage

No way this car is making this turn!” I think as I feel the car screechingly lose traction and drift rightward, the rest of the car trying and failing to follow the front wheels left.

“Bryan!” Nathan interjects from the back.

Galusha grabs the wheel, and just as the car starts to slide sideways off the track, he doesn’t commit the amateur move of doubling down on a bad turn and steering even further into it (thereby rolling the car) but instead straightens out. We shoot directly into the waterlogged patch of grass beyond the turn at around 30 mph, and just as the car bogs down in the grassy muck, cool-headed Galusha gives it a bit of gas to maintain momentum and brings us slowly back alongside pavement. There’s a pause as everyone catches their breath and regains normal consciousness again.

“That only happened because there’s a journalist here. It’s like demo gremlins,” I crack, trying to lighten the mood. Not many words are exchanged as the human, not the code, drives us back to the pits so we can see what went wrong.

"Throttle … OK, brake … Slight throttle … Full throttle."

Racing coach Gregg Franklin is sitting in the passenger seat, and on account of the earplugs, helmets, and engine noise, every order is accompanied by a hand signal in my peripheral vision—bunched fist for brakes, a forward flat-hand salute for throttle—telling me which pedal to hit.

The author tried his hand at driving the track, piloting his own Ford Focus RS. The goal: beat the robot cars in a time trial.

David Robert Elliott

This time the driver is me, and I’m struggling to get my car, a 2017 Ford Focus RS (the racy version of the family hatchback), down the same racetrack without the benefit of precision GPS. A tenth of a meter? I’m not sure I’m even within the same zip code as the racing line, and Franklin constantly corrects my approaches to the course’s turns.

He also constantly gestures for more throttle, which I hit on sheer faith alone, as I’m not sure even the car’s sporty suspension and all-wheel drive will carry it around another screeching gyration.

Track racing involves alternately standing on the gas or the brake pedal the entire time, until your car reeks of brake and clutch and you’re covered in sweat and mentally exhausted. I’ve definitely reached that point when the checkered flag signals all of us to head back into the pits. On Franklin’s instructions, I do a few circles in the pits to let the brakes cool down, and then I get out and walk around to let myself do the same.

In the pits, the various teams continue futzing with the cars, oblivious to the various human-driven cars roaring away on the track. Many of the autonomous-vehicle founders and employees are (unsurprisingly) car enthusiasts themselves, so they’ve been sneaking a few laps with their own modded-out cars in between autonomous runs (as I had).

The vibe puts me in mind of the Homebrew Computer Club—the storied meetup where the two Steves (Jobs and Wozniak) first showed off the Apple I: informal, improvisational, and filled with a mix of hobbyists and entrepreneurs messing around in the early days of a transformative technology.

One car, a Kia Niro hybrid festooned with logos from another startup, PolySync, is being worked on by two different teams, each on one end of the car. A team from online learning startup Udacity, which offers a “nano” degree in autonomous-car engineering, is affixing a camera to the windshield.

A small gadget, called an inertial measurement unit (which is like the accelerometer in your phone) sits on its official mounting bracket—a coffee cup—in the center console. Back in the trunk, the founder of another startup called Qibus is messing with the remote driving technology they are developing, which allows you to drive a car anywhere from your laptop.

An inertial measurement unit—a device akin to the accelerometer in your phone—can detect a vehicle’s motion, whether it’s sliding, and its orientation with respect to the road.

David Robert Elliott

The ability for two teams to work on the same car is the point of PolySync’s technology (as well as a growing part of the autonomous world). Nestled under the driver’s seat is its DriveKit, a blinking box that takes electronic inputs like steering angle or brake and throttle, and translates them into actual movements of the car’s controls.

The fly-by-wire systems of modern cars mean that you no longer need actual mechanical force to drive a car: The feel in the steering wheel that makes you think you’re mechanically coupled to the front wheels (which you are, just barely) is a mostly artificial thing controllable via software. You really just need the right signals into the right wires to make the car do anything.

To prove the point, PolySync CTO Ken Keiter plugs an Xbox game controller into a laptop, connects that to his blinking box, and offers to take me around the track. Getting into the passenger seat for additional lulz, Keiter proceeds to steer the car from there using the game controller: left trigger button for brake, right trigger for throttle, and left-right on the joystick to move the wheel.

It’s a videogame come to life. With the right wireless technology (like the one in the trunk of the car), he could have been driving from the other side of the globe. Just another piece of the technological puzzle in this emerging industry.

Ken Keiter, the CTO of a startup called PolySync, operates a car using an Xbox game controller.

David Robert Elliott; Animation by WIRED Staff

Which is precisely the point. In the “mine for gold” vs. “sell picks and shovels to miners” trade-off of any booming startup ecosystem, many entrepreneurs opt to supply parts to companies toiling in the riskiest end of the business.

For another example of the same phenomenon, consider Warren Ahner, founder and CEO of Right Hook, which provides sophisticated testing and simulation software for autonomous vehicles. The company creates what’s effectively a Matrix for your driving algorithm: a simulated reality you plug into your software, allowing you to test it in the confines of code rather than on a busy urban street (or the Thunderhill raceway).

In addition to simulation, Right Hook’s algorithm allows fast-moving startups to perform so-called regression tests on their code. In a production environment, when engineers push a new code change, the software makes sure that prior logic stays correct: Critical mathematical calculations still need to spit out the right answers.

In the case of self-driving cars, that’s a vehicle not hitting a pedestrian on, say, a particularly complicated traffic circle. “You can see where your edge cases are, test them, and make them harder and harder as your code gets better and better,” Ahner says. The majors like Waymo and Uber test their vehicles over billions of miles of roadway, but the smaller companies can’t afford that, so Right Hook’s Matrix it is.

So: There are lots and lots of startups here, many of them either fielding cars or just nosing around the pits at the track, looking to fill in some blanks in the self-driving crossword.

The scene is like the Homebrew Computer Club, yes, but another historical parallel also presents itself. The early consumer internet was like the current batch of autonomous-vehicle hegemons. Today’s Waymo, Uber, and General Motors’ Cruise are reminiscent of incumbents like Compuserve and AOL: vertically integrated do-everything companies practically synonymous in users’ heads with their emerging industry. There might have been other technologies and protocols under the hood, so to speak, but the product presented itself as a titanic, unitary whole.

This turned out to be an illusion. The underlying modular nature of the internet soon revealed itself, with everyone and their sister biting off a piece of the integrated whole and then racing to seize horizontal market share for that slice of functionality, rather than assail the incumbents’ verticality.

Eventually, that AOL login refrain “You’ve got mail!” became nothing more than a Meg Ryan rom-com, as some companies seized the topmost layer that users see (Facebook), while others dominated the underlying plumbing (CloudFlare). The companies at the track event are effectively betting on a similar modularized collection of technologies eventually dominating the autonomous-vehicle biz, each of them owning some small layer of it.

This assumes, of course, that progress in self-driving cars, like that of the consumer internet, turns out to be valuable for an ever diversifying set of uses. In actual fact, everyone I speak to shares a muted despair at the slow progress that autonomous vehicles have made, not even remotely living up to their early promise. “This is the trough of disillusion­ment,” says PolySync’s Keiter, referring to the tech doldrums that often come between the precipitous hoopla of media coverage and funding rounds, and the eventual, slow upward progress of technology years later.

Insiders categorize a technology’s progress from level 1 autonomous driving—something basic like adaptive cruise control—to level 5, full Terminator-style automation where the software has total control. It’s clear that level 4 or 5 automation on city streets—the day your Uber driver disappears—isn’t happening anytime soon.

For now the progress is happening at narrower scales. For example, a company called Voyage is building fully autonomous cars for the controlled environment of retirement communities, those vast cities-within-cities of endless Spanish tile roofs and swimming pools in South Florida. By having set routes, changing on-the-ground infrastructure if need be, and (perhaps most importantly) providing remote human support to handle driving hiccups, the company hopes to provide the self-driving dream to retirees and expand from there. “Humans will be in the loop for a long, long time,” says Voyage CEO Oliver Cameron.

Human remote control is the tack taken by companies in another promising arena: long-haul trucking. Starsky Robotics is building a fleet of 18-wheelers that are semi-autonomous but remotely controlled by humans. Embark (which had a truck at the event—unfortunately, over the weight limit of the track) is building a fleet with level 2 autonomy that still features a human in the cab; one such vehicle recently completed a coast-to-coast trip.

In short, while much public attention is still focused on the futuristic city-driving dream, the real progress is often happening in transport niches largely out of view. One day, the technologies developed there will emerge, even if only as added safety features on (mostly) human-driven cars.

And how about in that other limited environment, somewhat more interesting than retirement communities—the racetrack?

A much bigger-budget version of the Thunderhill competition is getting underway with the unimaginatively named RoboRace, a racing concept using the same format as the electric-car Formula E series. Their first car is named (you guessed it) RoboCar and sports a pointed nose and flared, biplane-like fenders that vaguely resemble a Romulan warbird from Star Trek: The Next Generation.

Conspicuously absent is a cockpit; only a narrow, wasp-like backbone connects the front set of wheels to the rear one. The vehicle has completed closed-course, single-car time trials, but as with the homebrew cars at Thunderhill, even the big-ticket autonomous racing tech isn’t good enough for full-on, wheel-to-wheel racing alongside several cars. The next version of RoboRace’s car will feature a human safety driver inside, making it possible to put several cars on the track at once.

And my racing? How well did I manage against the various autonomous teams? Armed with good coaching and a speedy car but no actual track skills, I clocked an OK-but-not-amazing 1 minute, 48 seconds. The Point One team managed the fastest autonomous track time of 2 minutes, 6 seconds, a full 18 seconds slower than this human driver.

Unlike more cerebral arenas like chess or Go, humans still have the advantage behind the steering wheel—on the racetrack and on city streets. It was a milestone when IBM’s Deep Blue defeated chess master Gary Kasparov. But the machines have a long way to go before they can defeat us in driving contests of our own devising.


Read more: