In September, 2014, I attended an MIT conference that explored the major progress that’s taken place in artificial intelligence, robotics and related technologies over the past several years. Autonomous vehicles was one of the main areas of discussion. With most other topics, there was considerable consensus, but not so with self-driving cars. While some thought that fully autonomous vehicles will be all around us within a decade or so, others were not quite so sure, myself included, due to the many highly complex technical, economic and societal issues that must be worked out.
I was reminded of this meeting a few weeks ago when I read that a Florida man had been killed while driving a Model S Tesla in autopilot mode. The accident is still under investigation, but it appears that the Tesla struck a tractor-trailer truck that was making a left turn in front of its path. Neither the driver or the Tesla’s autopilot noticed that a truck was suddenly crossing its lane of traffic, perhaps because the white truck was hard to spot against a bright sky.
This accident has led to a renewed discussion of the current state-of-the-art of vehicle automation, the approaches being pursued by different companies, and the prospects for the near- and long-term future.
This is the case with the application of automation and AI to driving. The National Highway Traffic Safety Administration (NHTSA) has defined 5 distinct levels of vehicle automation:
- No-Automation (Level 0): The driver is in complete and sole control at all times.
- Function-specific Automation (Level 1): Driver has complete authority but is helped by automation in certain normal or crash potential situations, e.g., electronic stability control and anti-lock braking systems.
- Combined Function Automation (Level 2): Driver can cede active control to at least two primary functions but is still responsible for monitoring and safe operations and expected to be available at all times, e.g., adaptive cruise control and lane departure warning systems.
- Limited Self-Driving Automation (Level 3): Driver can cede full control under certain traffic and environmental conditions, but must be available at all times to take back controls when needed.
- Full Self-Driving Automation (Level 4): Human provides destination and/or navigation input but is not expected to be available to control the vehicle at any time during the trip. Vehicle is solely responsible for safe operations and could be traveling occupied or empty.
The auto industry has been adding level 1 features to their cars for over 40 years, and more sophisticated level 2 features for the past couple of decades. Generally, these are first introduced in higher end cars, but as their costs drop over time, the features are then available in lower priced cars as well. While some of the features are aimed at user convenience, safety is far and away the overriding objective.
The NHTSA estimates that there were over 35,000 traffic deaths in 2015, an increase of 7.7 percent over 2014 fatalities. In addition, over 2.3 million people were injured in 2014. 94 percent of crashes can be tied back to human error. Over the past decade, close to 100 people have been killed in traffic accidents each day in the US alone.
These figures are stark reminders that, as NHSTSA Administrator Mark Rosekind noted, “we need to focus our efforts on improving human behavior while promoting vehicle technology that not only protects people in crashes, but helps prevent crashes in the first place.” Vehicle automation, including autonomous vehicle AI research, is thus a very important priority to try to significantly reduce the 94 percent of fatal crashes involving human error.
Toyota, for example, recently established the Toyota Research Institute to pursue the development of what it calls guardian angel automation technologies, where the driver is always in control until a crash looms, at which point the automation system fully takes over to try to prevent the accident. Toyota is also pursuing the development of level 4 self-driving cars, but its guardian angel approach appears more promising for the near future. It’s clearly positioned as an advanced level 2 capability where the driver is always in charge unless an accident is imminent. Its key objective is saving human drivers from their all-too-frequent errors, - not replacing them.
While just about all car companies continue to seriously invest in advanced level 1 and 2 technologies, their approaches toward level 4 are quite different. Some, like Tesla, have taken a more incremental level 3 approach, which allows the driver to cede control of the car under certain conditions while emphasizing that drivers must remain alert, engaged and always ready to take over. But, as a number of recent articles have pointed out, the autopilot feature is so compelling that drivers soon get comfortable and are lulled into feeling that they can turn their attention away from the road. Then can then get distracted with other activities and take their hands off the wheel, no matter how much they’ve been warned not to do so.
“Experiments conducted last year by Virginia Tech researchers and supported by the national safety administration found that it took drivers of Level 3 cars an average of 17 seconds to respond to takeover requests,” noted a recent NY Times article. “In that period, a vehicle going 65 m.p.h. would have traveled 1,621 feet - more than five football fields.”
According to this recent LA Times article, Google chose a different path from Tesla based on its early experiences with its self-driving car prototypes. “Once behind the wheel of the modified Lexus SUVs, the drivers quickly started rummaging through their bags, fiddling with their phones and taking their hands off the wheel - all while traveling on a freeway at 60 mph.” In a panel earlier this year, Chris Urmson, - head of Google’s self-driver program, - noted that “Within about five minutes, everybody thought the car worked well, and after that, they just trusted it to work. It got to the point where people were doing ridiculous things in the car.”
“After seeing how people misused its technology despite warnings to pay attention to the road, Google has opted to tinker with its algorithms until they are human-proof… focusing on fully autonomous vehicles - cars that drive on their own without any human intervention and, for now, operate only under the oversight of Google experts.”
But, fully autonomous cars still face many challenges. For example, Gill Pratt, who was recruited from DARPA to lead the Toyota Research Institute, noted earlier this year that a fully autonomous vehicle must be able to handle highly unusual situations it had never encountered before, such as avoiding a mattress falling off a moving truck on a crowded highway. Handling such challenges will take time.
Given that fully autonomous level 4 vehicles are likely many years in the future, some car companies believe that the more incremental transition to level 3 is still a good intermediate step to take, especially if it’s done under highly specific circumstances aimed at reducing overall traffic accidents. Audi, for example, is planning to introduce level 3 features for some of its cars in the near future, such as stop-and-go traffic on the highway at no more than 35 mph.
Finally, there are other innovations to be explored beyond the automation levels defined by the NHTSA. While autonomous machines find it very hard to operate in highly unpredictable environments, we’ve long been adapting and simplifying environments so we can benefit from what machines are good at. In highly selected environments, such as moving between terminals in an airport, the trains can be fully automated and not require a human operator.
Such approaches would enable us to co-design our vehicles along with the environment in which they will operate. We could imagine the development of cars and trucks that can only go into self-driving mode when they’re in specially instrumented traffic lanes limited to such vehicles. Once outside those lanes, the vehicles would revert to their more classic human-controlled mode. Such hybrid approaches to vehicle automation might be a reasonable intermediate step on our way to full vehicle automation.
The AI research going into vehicle automation is very important because the stakes are so high, - saving many, many thousands of lives around the world. Most of the technologies being developed will significantly improve the overall safety of our cars and help reduce our large numbers of traffic accidents, deaths and serious injuries. And perhaps at some point in the future we will see self-driven vehicles coursing along our streets and highways.