From Silicon To Software


Where’s My Autonomous Vehicle?

To better understand the challenges ahead for fully autonomous vehicles, research teams over the last few decades have attempted to automate the process of driving. But early successes have not yet given us truly autonomous vehicles. Why?

The Defense Advanced Research Projects Agency (DARPA) created the first autonomous vehicle in 1984. This limited-use autonomous vehicle could drive on- and off-road; however, it was limited in range. Subsequent successes in increasing the range and robustness of the vehicle suggested the age of the fully autonomous vehicle was possible soon.

Early successes were so promising that Sergey Brin of Google announced in 2012 that his company would provide self-driving cars by 2017. Volvo, in 2014, announced that it would have a “Drive Me” program up and running in Sweden by 2017 as well. And in 2016, the Ford Motor company announced it would have production-grade self-driving cars on the market by 2021. Those predictions either did not come true or have since been rescheduled out a few more years because early successes with autonomous driving have created more questions than answers.

To assist with learning, DARPA in 2004 funded a competition to award a cash prize to a college engineering team that built a vehicle that could successfully navigate an off-road course in the Mohave Desert. This was an ambitious 150-mile route along Interstate 15 from Barstow, California to Primm, California. Perhaps not too surprisingly, none of the vehicles completed the route.

That didn’t, however, stop the challenge.

In 2005 teams competed once again in another DARPA Grand Challenge. For this, an autonomous vehicle had to drive only 17 miles; however, it was complicated with three tunnels and a cliff’s edge. Again, most of the vehicles didn’t complete the entire course. Four did, and each did so in less than eight hours.

What’s remarkable is that none of these vehicles were originally designed to operate autonomously. They were existing vehicles kludged with sensors, cameras, and transmitters strapped to all sides and the top. Remote team members then huddled nearby, like NASA scientists watching the monitors to see what a rover was doing on another world.

If driving the open road had shown some success, what about a more complicated scenario? For that, the DARPA Grand Challenge switched to a mock urban environment in 2007. Instead of cliffs with sheer drop-offs, the vehicles had to navigate a closed 60-mile course with real-time variables (like other vehicles and pedestrians) as well as follow all traffic rules. The vehicle teams were divided into two groups so their paths would intersect at various points along the way.

The winning engineering teams were from Stanford, Carnegie Mellon, and Virginia Tech. Each relied upon augmented off-the-lot Chevy Tahoes, Ford Escapes, and Volkswagen Passats. For the urban challenge, there also needed to be modifications to the internal automotive software. The winning software code was later released to the public as open source software.

In 2010, engineers at Volkswagen independently took the autonomous vehicle challenge to new heights by having a 2010 VW Touareg scale Pike’s Peak in Colorado, a mountain that rises 4,721 feet at an average grade of 7 percent. According to Wired, the vehicle was chosen because it featured a fly-by-wire throttle, adaptive cruise control, and a semiautomatic DSG gearbox. It also had the Oracle Java Real-Time Operating System, Oracle Solaris, and a built-in GPS, which engineers augmented with advanced algorithms. During one week of testing, the VW made five successful runs.

But that was almost ten years ago. What about today?

The autonomous vehicles being tested today are still modified production vehicles. Google’s Waymo division does have a concept vehicle it built itself, but it is unclear whether the company plans to mass produce it in the future. And these self-driving vehicles are being tested in urban areas with strong connectivity, not out on the open roads.

As of November 2017, Google racked up over four million miles of data from driving their autonomous vehicles on public roads. They have worked through more than 20,000 unique scenarios including “jumping out of canvas bags or skateboarders lying on their boards,” with more to come. In comparison, Uber has crossed one million road miles with its autonomous vehicles.

As impressive as that sounds, these tests have largely been conducted on the West Coast, in favorable weather, under controlled circumstances. What happens when Lidar, sonar, or radar sensors are caked with snow, or when temperatures drop below zero degrees Fahrenheit? More learning is needed, as more robust chips need to be designed and more advanced software needs to be created before we get to the age of the fully autonomous vehicle.