DARPA Grand Challenge: 20 Years Later

The 2004 DARPA Grand Problem was a spectacular failure. The Protection Superior Analysis Tasks Company had supplied a US $1 million prize for the group that might design an autonomous floor car able to finishing an off-road course by way of generally flat, generally winding and mountainous desert terrain. As IEEE Spectrumreported at the time, it was “the motleyest assortment of autos assembled in a single place for the reason that filming of Mad Max 2: The Highway Warrior.” Not a single entrant made it throughout the end line. Some didn’t make it out of the parking zone.

Movies of the makes an attempt are comical, though any laughter comes on the expense of the various engineers who spent numerous hours and tens of millions of {dollars} to get to that time.

So it’s all of the extra outstanding that within the second DARPA Grand Challenge, only a yr and a half later, 5 autos crossed the end line. Stanley, developed by the Stanford Racing Team, eked out a first-place win to assert the $2 million purse. This modified Volkswagen Touareg [shown at top] accomplished the 212-kilometer course in 6 hours, 54 minutes. Carnegie Mellon’s Sandstorm and H1ghlander took second and third place, respectively, with instances of seven:05 and seven:14.

Kat-5, sponsored by the Grey Insurance coverage Co. of Metairie, La., got here in fourth with a decent 7:30. The car was named after Hurricane Katrina, which had simply pummeled the Gulf Coast a month and a half earlier. Oshkosh Truck’s TerraMax additionally completed the circuit, though its time of 12:51 exceeded the 10-hour time restrict set by DARPA.

So how did the Grand Problem go from a complete bust to having 5 strong finishers in such a brief time frame? It’s positively a testomony to what could be achieved when engineers rise to a problem. However the final result of this one race was preceded by a for much longer path of analysis, and that plus just a little little bit of luck are what finally led to victory.

Earlier than Stanley, there was Minerva

Let’s again as much as 1998, when pc scientist Sebastian Thrun was working at Carnegie Mellon and experimenting with a really totally different robotic: a museum tour information. For 2 weeks in the summertime, Minerva, which seemed a bit like a Dalek from “Physician Who,” navigated an exhibit on the Smithsonian National Museum of American History. Its most important process was to roll round and dispense nuggets of details about the displays.

Minerva was a museum tour-guide robotic developed by Sebastian Thrun.

In an interview on the time, Thrun acknowledged that Minerva was there to entertain. However Minerva wasn’t only a folks pleaser ; it was additionally a machine learning experiment. It needed to be taught the place it might safely maneuver with out taking out a customer or a priceless artifact. Customer, nonvisitor; show case, not-display case; open ground, not-open ground. It needed to react to people crossing in entrance of it in unpredictable methods. It needed to be taught to “see.”

Quick-forward 5 years: Thrun transferred to Stanford in July 2003. Impressed by the primary Grand Problem, he organized the Stanford Racing Workforce with the purpose of fielding a robotic automotive within the second competitors.

In an enormous oversimplification of Stanley’s most important process, the autonomous robotic needed to differentiate between highway and not-road with a purpose to navigate the route efficiently. The Stanford group determined to focus its efforts on growing software program and used as a lot off-the-shelf {hardware} as they might, together with a laser to scan the quick terrain and a easy video digital camera to scan the horizon. Software program overlapped the 2 inputs, tailored to the altering highway circumstances on the fly, and decided a protected driving velocity. (For extra technical particulars on Stanley, try the team’s paper.) A remote-control kill switch, which DARPA required on all autos, would deactivate the automotive earlier than it might change into a hazard. About 100,000 strains of code did that and rather more.

The Stanford group hadn’t entered the 2004 Grand Problem and wasn’t anticipated to win the 2005 race. Carnegie Mellon, in the meantime, had two entries—a modified 1986 Humvee and a modified 1999 Hummer—and was the clear favourite. Within the 2004 race, CMU’s Sandstorm had gone furthest, finishing 12 km. For the second race, CMU introduced an improved Sandstorm in addition to a brand new car, H1ghlander.

Most of the different 2004 rivals regrouped to strive once more, and new ones entered the fray. In all, 195 groups utilized to compete within the 2005 occasion. Groups included college students, teachers, trade specialists, and hobbyists.

After web site visits within the spring, 43 groups made it to the qualifying occasion, held 27 September by way of 5 October on the California Speedway, in Fontana. Every car took 4 runs by way of the course, navigating by way of checkpoints and avoiding obstacles. A complete of 23 groups had been chosen to try the principle course throughout the Mojave Desert. Competing was a pricey endeavor—CMU’s Pink Workforce spent greater than $3 million in its first yr—and the names of sponsors had been splashed throughout the autos just like the logos on race automobiles.

Within the early hours of 8 October, the finalists gathered for the massive race. Every group had a staggered begin time to assist keep away from congestion alongside the route. About two hours earlier than a group’s begin, DARPA gave them a CD containing roughly 3,000 GPS coordinates representing the course. As soon as the group hit go, it was palms off: The automotive needed to drive itself with none human intervention. PBS’s NOVA produced a superb episode on the 2004 and 2005 Grand Challenges that I extremely advocate if you wish to get a really feel for the joy, anticipation, disappointment, and triumph.

Within the 2005 Grand Problem, Carnegie Mellon College’s H1ghlander was certainly one of 5 autonomous cars to complete the race.Damian Dovarganes/AP

H1ghlander held the pole place, having positioned first within the qualifying rounds, adopted by Stanley and Sandstorm. H1ghlander pulled forward early and shortly had a considerable lead. That’s the place luck, or fairly the shortage of it, got here in.

About two hours into the race, H1ghlander slowed down and began rolling backward down a hill. Though it will definitely resumed transferring ahead, it by no means regained its high velocity, even on lengthy, straight, stage sections of the course. The slower however steadier Stanley caught as much as H1ghlander on the 163-km (101.5-mile) marker, handed it, and by no means let go of the lead.

What went fallacious with H1ghlander remained a thriller, even after intensive postrace evaluation. It wasn’t till 12 years after the race—and as soon as once more confidently—that CMU found the issue: Urgent on a small digital filter between the engine management module and the gas injector precipitated the engine to lose energy and even flip off. Workforce members speculated that an accident a couple of weeks earlier than the competitors had broken the filter. (To be taught extra about how CMU lastly figured this out, see Spectrum Senior Editor Evan Ackerman’s 2017 story.)

The Legacy of the DARPA Grand Problem

No matter who gained the Grand Problem, many success tales got here out of the competition. A yr and a half after the race, Thrun had already made nice progress on adaptive cruise control and lane-keeping help, which is now available on many industrial autos. He then labored on Google’s Avenue View and its preliminary self-driving cars. CMU’s Pink Workforce labored with NASA to develop rovers for doubtlessly exploring the moon or distant planets. Nearer to residence, they helped develop self-propelled harvesters for the agricultural sector.

Stanford group chief Sebastian Thrun holds a $2 million test, the prize for profitable the 2005 Grand Problem.Damian Dovarganes/AP

After all, there was additionally lots of hype, which tended to overshadow the race’s militaristic origins—bear in mind, the “D” in DARPA stands for “protection.” Again in 2000, a defense authorization bill had stipulated that one-third of the U.S. floor fight autos be “unmanned” by 2015, and DARPA conceived of the Grand Problem to spur improvement of those autonomous vehicles. The U.S. military was nonetheless fighting in the Middle East, and DARPA promoters believed self-driving autos would assist decrease casualties, notably these attributable to improvised explosive units.

DARPA sponsored extra contests, such because the 2007 Urban Challenge, during which autos navigated a simulated metropolis and suburban surroundings; the 2012 Robotics Challenge for disaster-response robots; and the 2022 Subterranean Challenge for—you guessed it—robots that might get round underground. Regardless of the competitions, continued navy conflicts, and hefty authorities contracts, precise advances in autonomous navy autos and robots didn’t take off to the extent desired. As of 2023, robotic floor autos made up solely 3 p.c of the worldwide armored-vehicle market.

In the present day, there are only a few absolutely autonomous floor autos within the U.S. navy; as an alternative, the companies have solid forward with semiautonomous, operator-assisted programs, akin to remote-controlled drones and ship autopilots. The one Grand Problem finisher that continued to work for the U.S. navy was Oshkosh Truck, the Wisconsin-based sponsor of the TerraMax. The corporate demonstrated a palletized loading system to move cargo in unmanned vehicles for the U.S. Military.

A lot of the up to date reporting on the Grand Problem predicted that self-driving automobiles would take us nearer to a “Jetsons” future, with a self-driving car to ferry you round. However 20 years after Stanley, the rollout of civilian autonomous automobiles has been confined to particular purposes, akin to Waymo robotaxis transporting folks round San Francisco or the GrubHub Starships struggling to ship meals throughout my campus on the College of South Carolina.

I’ll be watching to see how the expertise evolves outdoors of massive cities. Self-driving autos could be nice for lengthy distances on empty nation roads, however elements of rural America nonetheless wrestle to get enough cellphone protection. Will small cities and the areas that encompass them have the bandwidth to accommodate autonomous autos? As a lot as I’d wish to suppose self-driving autos are almost right here, I don’t anticipate finding one below my carport anytime quickly.

A Story of Two Stanleys

Not lengthy after the 2005 race, Stanley was able to retire. Recalling his expertise testing Minerva on the Nationwide Museum of American Historical past, Thrun thought the museum would make a pleasant residence. He loaned it to the museum in 2006, and since 2008 it has resided completely within the museum’s collections, alongside different outstanding specimens in robotics and automobiles. The truth is, it isn’t even the primary Stanley within the assortment.

Stanley now resides within the collections of the Smithsonian Establishment’s Nationwide Museum of American Historical past, which additionally homes one other Stanley—this 1910 Stanley Runabout. Behring Heart/Nationwide Museum of American Historical past/Smithsonian Establishment

That distinction belongs to a 1910 Stanley Runabout, an early steam-powered automotive launched at a time when it wasn’t but clear that the internal-combustion engine was the way in which to go. Regardless of clear drawbacks—steam engines had a nasty tendency to blow up—“Stanley steamers” had been recognized for his or her wonderful craftsmanship. Fred Marriott set the land speed record whereas driving a Stanley in 1906. It clocked in at 205.5 kilometers per hour, which was considerably sooner than the Twenty first-century Stanley’s common velocity of 30.7 km/hr. To be truthful, Marriott’s Stanley was racing over a flat, straight course fairly than the off-road terrain navigated by Thrun’s Stanley.

Throughout the century that separates the 2 Stanleys, it’s simple to hint a story of progress. Each are clearly recognizable as four-wheeled land autos, however I believe the science-fiction dreamers of the early twentieth century would have been hard-pressed to think about the suite of applied sciences that may propel a Twenty first-century self-driving automotive. What’s going to the autos of the early twenty second century be like? Will they even have 4 tires, or will they run on one thing fully new?

A part of a continuing series historic artifacts that embrace the boundless potential of expertise.