Never miss a story from Breakfast Bytes. Subscribe for in-depth analysis and articles.
Recently Tesla activated Autopilot for their vehicles, which are equipped with the appropriate sensors. To prepare for this addition, they have been shipping cars and SUVs for about a year with front-facing camera and radar and a collection of sensors. Autopilot allows the vehicle to follow the lanes, follow cars in front safely, change lanes, and park itself. It is not a truly autonomous vehicle since it doesn't understand street signs and all traffic regulations. Elon Musk in a recent analyst call said that he thought true autonomy was about three years away. So autonomous vehicles don't look to be that far off in the future.
But as recently as ten years ago it didn't look that way.
In 2004, DARPA organized the "Grand Challenge" for driverless cars, offering a prize of $1M. It was held on a 120-mile route in the Mojave desert. That turned out to be a lot of unnecessary advance planning since the furthest that any vehicle got was 7.3 miles by Carnegie Mellon University's (CMU's) Sandstorm. The race was a fiasco with a couple of the 21 qualifying vehicles failing to make the start. One flipped in the starting area (although to be fair, it was a self-driving motorbike), several crashed after just a few hundred feet before or at the first turn. 95% of the course was unused. The $1M prize was not awarded, it was rolled forward to the following year. Watch the video below.
In 2005, things had improved by an almost unbelievable amount.There was a $2M prize and a new course involving 100 turns, three tunnels, and, at the end, a pass with a steep drop on one side and a cliff on the other. There were 23 finallists and all but one got further than the 7.3 miles of the previous year's winner. Five vehicles finished the entire 132 mile course. The winner, based on time, was Stanford's Stanley in 6 hours 54 minutes. Second was one of CMU's entries, H1lander, 11 minutes slower, and third was the 2004 winner, CMU's Sandtorm 9 minutes after that. See the video below.
There was no challenge in 2006. In 2007, the so-called "Urban Challenge" was held at a disused air force base in Southern California. Unlike previous years in the desert, vehicles had to obey traffic laws, avoid other vehicles, and merge with traffic. The other vehicles on the course were the other competitors and professional drivers. The course was 60 miles long. The $2M winner was Tartan Racing, a collaborative effort by CMU and GM, with a modified Chevy Tahoe called Boss. Second, winning $1M, was the Stanford racing team with an adapted VW Passat, Junior. Third, winning $500K was Virginia tech's Ford Escape, Odin. See the video below.
The big-picture view is that, 11 years ago, the furthest a self-driving vehicle could go was under 10 miles on a desert course cleared of other vehicles. Just four years later, many vehicles finished an urban driving course with other vehicles, four-way stops, and more. It is an incredible advance in a short period of time. After that, attention switched from the DARPA-sponsored challenges to private-sector developments.
The most well-known, at least if you live in the Bay Area, is the Google self-driving car, which is a common sight on the freeways around here with its distinctive spinning LIDAR on the roof (but not to be confused with Google Street View mapping vehicles that also have strange camera apparatus on the roof). The Google self-driving car project was headed up by Sebastian Thrum who led the Stanford team that won the 2005 challenge and came second in 2007 (urban).
At DAC this year, Jeffrey Owens, the CTO of Delphi (originally a division of GM making electronic components, but spun out as a separate company) talked about the Delphi-equipped Audi that had driven from San Francisco to New York. If you were at DAC, you probably saw the vehicle itself on the show floor. Jeffrey pointed out that the sensors were integrated into the body work or behind the rear-view mirror. He said that aesthetics are very important and a car with a huge spinning LIDAR on the roof will not sell. When a minivan can fail to sell due to having too few cupholders, making the technology unobtrusive will be very important.