I am ambivalent about the current efforts being expended by Google and others on autonomous vehicles. As a society we appear to have backed ourselves into a corner where the only way out is to shift a ton of metal alongside driver and passengers. Self-driving cars don’t do much to change this, and hand-wringing articles about how the trolley problem might be resolved by autonomous vehicles are a distraction.
Nonetheless, it appears that the efforts of Google and others will produce self-driving vehicles for the mass market, so it is interesting to look at the work they are doing. Google has been particularly forthcoming about their research, as shown in this TED talk by Chris Urmson, who has headed up Google’s self-driving car programme since 2009.
There is a lot of interesting information here about what Google is doing with its cars and how they are coping with real-world traffic situations. He is particularly persuasive on the safety point — human beings are responsible for many more accidents resulting in death or serious injury. If the future of transportation has to involve cars, far better that those cars are not driven by distracted and borderline incompetent human beings.
But the point I found most intriguing comes when we are introduced to the way Google’s car sees the world. (This runs from about 7’48” in the video.)
The starting point, basic driving on grade-separated highways, centres on perception (where the car sees itself and the other road-users in the world) and experience (what has happened before that might happen again). This is roughly where Google were when they started serious work in 2009. (What Chris Urmson calls “a geometric view of the world.”)
Once Google moved the cars onto city streets, there was immediately much more complex information to handle. The cars needed to be aware of objects other than other vehicles — pedestrians, animals, road works, litter, and so on. At this point, the cars need to be able to deal with a range of different signals — flashing lights on police vehicles or school buses, for example. They also need to judge and work around the behavioural expectations of other road users at a host of different levels — some signalled and some implicit. At 10’09”, Urmson tells us the key to their success in this effort:
The way we accomplish this is by sharing data between the vehicles.
At first this is just sharing information about the location of hazards like road works between vehicles, so that their shared understanding of the environment is constantly updated. Over time, this has developed into a massive shared database of all of the things that all of the cars have seen over time. Hundreds of thousands of objects have been observed by the cars in multiple dimensions. All these objects (cars, people, animals, cars, trucks, cyclists…) can be used by Google’s cars later on to help them understand novel objects and situations by comparison with what has already been seen and recorded. Further than that, this data can be used to build a model of how different objects might behave in the world — improving the predictive capability of the fleet immeasurably.
Google’s cars will always be better than humans in terms of their capacity for observation, speed of reaction and ability to deal with crises calmly and decisively. Urmson shows in detail how this is achieved even when a hazard is partly concealed by other traffic in the section of video from 12’30” to 13’24”. That gives them an edge as individuals. Their constant sharing of data is where the real differentiation occurs. They are constantly learning by sharing.
Humans will never have the processing capabilities of an autonomous car. But there will always be somethings that we can do immeasurably better than technology. The key to future success is working out what those things are and concentrating on them. But we can also learn from Google’s cars. By sharing what we know as widely as possible and actively using what is shared with us, we can develop a better picture of the world and act within it.
Your knowledge-sharing capability is almost certainly nowhere near as good as a dumb car. But the car can show you why you should be better at it. Google’s cars understand that more can be achieved by sharing than by hoarding. We should learn the same.