How close are human drivers to interacting with autonomous vehicles on the roads? That is the answer a panel at the North American International Auto Show in Detroit tried to solve on Wednesday. It didn’t come up with a definitive answer, and in fact, may have created more questions.
“If I look at where we are as an [autonomous vehicle [AV] community, we’re about 100 times worse than the best drivers in the world, and that’s U.S. drivers,” explained Shane Elwart, deputy chief engineer for autonomous driving systems for American Haval Motor Technologies, the U.S. subsidiary of China-based Great Wall Motors.
Elwart said that initial analysis of autonomous vehicle “incidents” shows that AVs are involved in 0.02 incidents – defined as any interaction with another vehicle, object or person regardless of injury – per 1,000 kilometers. The data doesn’t indicate how many of those incidents are due to the AV.
“That would result in over 200,000 incidents per billion miles if we turned that AV loose right now,” he said. Conversely, in the U.S., there were five million incidents in 2008 for a rate of 286 per billion miles. “And that’s the best of the best drivers,” Elwart added.
Before the audience could shut down the autonomous vehicle, though, Elwart noted the process is still early but industry needed to set verifiable goals. “What is the metric,” he asked. “How do we know when we are done [developing them]?”
He suggested industry adopt a 200 incident per billion miles goal. “We need to look across the entire [AV] industry and put out a metric that we can all meet,” Elwart noted. “If we are able to reduce this down to 200 incidents per billion miles, we would prevent two million incidents a year and save over 27,000 lives.”
Phil Magney, founder and principal of VSI Labs, which provides research and supports AV technologies, said he sees three trajectories going on right now in AV development: incremental deployment of Level 1 and 2 vehicles; development of “robo taxis” at the Level 4 level; and lower speed fixed route shuttles.
“OEMs are, of course, sometimes working on multiple tracks at the same time,” he said. “When you go beyond that you have trucking and mining and a lot of other areas” where autonomous vehicles are in development.
Several panelists brought expertise in specific areas of autonomous vehicle development. Chris Posch, director of OEM Tech Services for camera maker FLIR, said that the increased usage of thermal cameras could aid in reliability and safety of the vehicles.
FLIR’s thermal cameras can identify objects long before traditional visible cameras, which are being used on most AVs currently. He said that 75% of all pedestrians killed by cars are done so at night. A thermal camera can detect an object up to four times farther than a headlight, likely preventing many of these incidents.
“We’re able to see not only the pedestrians, but also whether they are riding a bike or not,” Posch said, while playing a video that showed just that. “Particularly with AEB (automatic emergency braking) with current systems using visible [cameras] and Lidar, you could be much safer by adding thermal cameras.”
Another issue holding back autonomous development is mapping issues. An AV needs strong mapping, which is something the nonprofit organization SpaceNet is working on. Ryan Lewis, vice president of In-Q-Tel CosmiQ Works, is involved in the SpaceNet initiative and explained its purpose.
“We use open-source remote sensor data focused on foundational mapping challenges,” he said. “We view remote sensing as having a very important role, particularly with base maps.”
Base maps are what AVs use to know where roads are, but Lewis said they need to do a lot more.
“Most of the remote sensing work today for base maps is based on [image] pixel count, and whether it looks like a road or is not a road,” he noted. “The key is not pixel count, but the connectivity of the road [to its surroundings].”
SpaceNet regularly conducts competitions to try and solve these problems, and Lewis said they have seen some algorithms that have done extremely well mapping areas, but others that have struggled to do so, particularly in heavily congested areas.
The panelists agreed that society, along with the technology of autonomous vehicles, is not ready yet. “Are we there yet?” Magney asked. “It depends on how you define there. Level 2 and 2-plus is alive and well, and that’s going to be huge to the [development and acceptance] over the next 10 years.”
He predicted that in the next few years, more deployment of autonomous vehicles in ride-sharing settings would begin, initially in very confined ways.
“It’s truly sobering to look at the numbers and wonder if we will ever get there,” Elwart concluded. “We need to find partners that really want to work on this.”