by Bill Cooke
Search engine giant Google is looking for partners within in the auto industry to help launch one of the most significant applications of artificial technology over the next several years, the self-driving car.
In a keynote address to the SAE 2012 World Congress on 25 April 2012, Anthony Levandowski, Business Lead for Google’s Self Driving Car Project provided an overview of Google’s autonomous vehicle program and requested that the auto industry partner with Google on the implementation. (Levandowski joined Google in 2007 to launch StreetView—Google Maps with Street View lets you explore places around the world through 360-degree street-level imagery.)
We’re not perfect; the technology is nowhere near ready. We want to set expectations low but we want to encourage dialogue on how we want to move the technology forward.—Anthony Levandowski
|“For some, driving is a distraction.” |
—Allen Taub, former GM VP, Global R&D
Levandowski shared that 32,788 people were killed in the US last year in auto accidents and 90% of those accidents were related to human error. Multi-tasking while driving is only increasing to the extent that people view driving as the distraction. Twenty percent of the food consumed in America is eaten in cars. Google believes that a future state with having computers drive cars can “remove a gigantic chunk” of the US fatalities.
Approximately 1.5 million people/year are killed in auto accidents globally. Google is involved because the company has a strong technical legacy and the company likes to take on problems where the “solutions have a high impact on humanity that involve challenging technical problems”.
In addition to the safety impact, Google believes your brain should be able to engage in activities other than driving.
It is a bug, not a feature, that you need to drive all of the time…What if I gave you a pill that allows you to get 10% longer life without any side effects …given how much time we spend in a car, a self driving car is that pill.—Anthony Levandowski
Prior to working for Google, many of the engineers were veterans of the 2004 and 2005 DARPA Grand Challenge that featured vehicles executing a route across an off road area. Many of these engineers also participated in the 2007 DARPA Urban Challenge which involved vehicles operating in an urban environment with defined routes but various fixed and moving hazards.
In 2009, Levandowski used an autonomous vehicle to deliver a pizza through the streets of San Francisco. It is interesting to note that in the interest of safety, and perhaps publicity, the pizza delivery vehicle was accompanied by an escort involving eight police cruisers and eight police motorcycles.
|“In God we trust, everyone else brings data”|
—Google design philosophy
Here’s the high-level approach to Google’s autonomous technology. When the vehicle is initialized and given a destination, it uses GPS to determine its position within 5-10 centimeters. It than uses Google’s mapping functions to overlay a model of the world and to create a route similar to an automotive GPS system. It uses a laser scanner, camera and radar to create a 3D model of the world around it and uses software to identify, categorize and track fixed and moving objects. It combines data from the GPS system and the real world measurements to create a path for the vehicle. The software has the ability to “anticipate” an open path but in general “the vehicle is more polite than you or I” according to Mr. Levandowski. The system focuses on defensive driving—“we don’t try to correct other people’s mistakes”—and it won’t rapidly accelerate to avoid being rear ended.
One of the advantages of the system is that it is backward compatible with existing infrastructure; it doesn’t require special markers in the road or transponders on other vehicles.
System performance isn’t limited to human reaction times. “We measure thousand of parameters during the vehicle development and one of the most significant data points is ‘time to collision’ (TTC).” TTC measures how quickly a collision is going to occur if no evasive actions are taken. In situations with a short TTC (less than 7.5 seconds) the robotic system is “dramatically better than a human.”
The vehicle is lightning fast through obstacle courses; “… over a closed course where a one minute lap time is considered excellent the robotic system can beat a professional driver by two to three seconds.”
Steve gets a taco and the hour-long commute. Google shared two examples of the development vehicle in use. The first one involved Steve, who is blind, and in 2012 got into a car and drove down to a local store to get a taco. This represented more of a future state since although the vehicle was capable of executing the actions to get Steve to and from the taco shop, there are still significant legal issues to be worked out before blind people can routinely drive cars autonomously. Google is interested in “Not only how we treat the ‘best’ but those that need our technology the most”.
The second example was a Google employee who lived in Berkeley, CA and commuted to Mountain View. The employee’s commute is approximately one hour each way and along that commute 45 minutes was in autonomous mode and 15 minutes was in driver mode with eight driver/autonomous transitions along the way. Per Google policy, employees weren’t allowed to text or do other highly distracting activities in autonomous mode but the employee really appreciated the ability to “concentrate on other things and make the mental transition between home and work” during autonomous mode.
The vehicle is capable of doing the entire commute autonomously but the team has decided to focus their efforts on improving highway performance. Google’s fleet of converted Priuses have travelled more than 250,000 miles in autonomous mode and Google believes they’ll need to covered closer to one million miles before the technology can be released onto the market.
From an energy standpoint, autonomous driving is more efficient since the vehicle is “hypermiling the entire time … and imagine if your vehicle knows the stop light ahead is going to be red before you get there …”
|“The real danger is the failure of our imagination.”|
Nevada and Florida (AB 511 and CS/HB 1207) have approved legislation that provides the legislative intent to safely develop the operation of motor vehicles with autonomous technology on the public roads of their respective states but doesn’t give authorization to deploy the vehicles until the technical and legal issues are better understood. The California state Senate has approved a similar measure (SB 1298).
The bill content varies by state but both states require a human to be in the vehicle during operation.
In Florida’s case, the legislation was approved by the Governor on 13 April 2012 and directs the State’s Department of Highway Safety for Motor Vehicles to prepare a report for the legislature on the safe operation of autonomous vehicles no later than 12 February 2014.
Nevada defines “Artificial intelligence” to mean the use of computers and related equipment to enable a machine to duplicate or mimic the behavior of human beings” and “autonomous vehicles” to mean a motor vehicle that uses artificial intelligence, sensors and global positioning system coordinates to drive itself without the active intervention of a human operator”.
Although the legal language is still very preliminary, Google hopes that ultimate legal responsibility and liability for the vehicle will stay with the owner but “as any third-grader will tell you, if the robot makes a mistake, it’s the robot’s fault”.
Failure modes and orange barrels. Google’s autonomous vehicles still have areas where they want to improve. Construction zones and joggers provide special challenges. Google is enthusiastically looking for worst case samples of each situation to develop their software. Levandowski showed a video of a narrow, winding CA road full of runners which they use for software development.
Google understands that there is always the risk of system failure and is especially concerned about a failure of a key sensor. There is going to be redundancy among the sensors but for some situations they want the driver to be able to take over in such a situation but they realize it can’t be instantaneous. They’re working on the appropriate window but they expect it to be on the order of 10-20 seconds.
|“There are as many views on this technology as there are car brands,” |
“Google doesn’t want to make cars” and is looking for partners to deploy the technology, Levandowski said. In addition to working with state legislators and NHSTA on policy and regulation they’ve also been in talk with insurance companies. Google believes they’ll be able to provide evidence that “actuaries can use to reduce insurance rates” to apply to vehicles while they’re being operated in autonomous mode.
Google has talked to various original equipment manufacturers (OEM) individually about the technology with various responses; Google has also been talking to the Auto Alliance—an industry wide group.
To accelerate the deployment of the technology, Google is exploring retrofitting existing vehicles. To prevent opposition from the OEM’s legal departments about third-party modifications, the Florida legislation explores provisions to protect the original equipment manufacturer from lawsuits if a third party converts their vehicle into autonomous drive.
Levandowski believes strongly in implementing the technology quickly: “the most important thing a computer can do over the next ten years is drive a car” and start reducing the 32,788 traffic fatalities a year in the US.