Autonomous Driving: Driving without Maps
Autonomous driving is difficult without detailed maps. Only a few companies have their own maps and enough staff to enter relevant information. Why today’s autonomous cars are not yet really autonomous and which developments lead there, is summarized here.
You can read it everywhere: in a few years autonomous vehicles will finally allow us to read AI blogs without danger while driving to work. But it's not that simple. We need to take a closer look at several levels to understand that it will probably take a little longer after all, since autonomous driving requires a major technological recirculation.
The question is what degree of driving ergonomics we are talking about. From simple assistance systems such as parking aid to level 5 autonomy (optional use of steering wheel and pedals), there are still some technical challenges to be solved. One of these challenges is mapping the world. The big players are competing with each other to create detailed, dynamic three-dimensional maps helping intelligent cars to find their way around the world better. But why do they need maps? Why are two-dimensional maps like Google Maps insufficient?
The answer lies in different error types that can occur during autonomous driving. Of course robotic cars can drive without maps and observe all relevant elements of the world themselves. But now and then they’d miss a red traffic light, cross a curb or a solid line or worst case crash into construction sites or buildings that would have been clearly marked on a map. Because every software, same as humans, makes a mistake at some point. In one of a thousand cases, for example, there is a notification missing (we call it "false-negative") and although this is an impressive technical achievement, it is not enough to entrust this system with a human life.
In order to reduce the probability of a serious accident, the system sensitivity is maximized - systems then react to the slightest indication of a traffic light. That way fewer traffic lights are missed, but there are also more error messages (false positives) if there is no traffic light nearby.
Nevertheless: better to stop once on a clear route than passing a red traffic light. However, with a map every detection can be checked for plausibility. The car can locate itself precisely by comparing its sensor data with the map. Maps reduce uncertainty.
Therefore almost all significant autonomous journeys so far have been carried out in mapped areas. And in this case, "mapped" does not only mean a simple one dimensional (flat) map. Most autonomous vehicles are equipped with many complex sensors that are ideal for mapping the world.
If a person follows a given route, you can look at the world from the car’s perspective and simply save all sensor data. This complex data material is subsequently processed and annotated by many screen workers; markings and meanings are added for example, so the autonomous vehicle knows exactly what to expect during the next run. This drastically reduces uncertainty operating an autonomous vehicle - but only if maps are up-to-date.
This leads us to the following prediction for the future of autonomous driving; although our cars are driving us, we won’t be able to let it drive anywhere by itself due to the fact that maps are not up to date. If necessary, we have to shift down a few autonomy levels and steer the car ourselves again.
Perhaps such a trip will also be reimbursed for bonus points, because once the car has safely passed through a blind spot, the valuable current data is only a mobile phone conversation away from the manufacturer or mapping service. Due to the huge amount of data this of course has implications for our data networks, a topic we will come back to.
Getting rid of maps
Now it seems to be logical to think about how to get rid of maps. They warrant a lot of manpower, deteriorate quickly and hinder our mobility if the car requires them. Recently Apple has filed a patent that deals with the above question. Unfortunately Apple has not given any details exactly how it can be done.
Strictly speaking, mechanical environment perception and correct interpretation of relevant conditions for navigation must improve (i.e. positioning of the car, positioning of other road users). This includes many levels - from evaluating individual sensors or the way they are merged to symbolic modelling of consequences of one's own actions.
Of course, this can not be covered in a single patent. But ideas that might lead to autonomy in broad market penetration in the coming decade have existed for a long time. So what are the next steps?
As we have learned in the past, the solution to a difficult task is not necessarily archived by improving the approach but in simplifying the problem. Let's take the traffic light as an example. Traffic lights have been introduced for human drivers to simplify one’s behaviour at intersections. This has reduced accident numbers as well as points of danger.
We will certainly see adjustments to our transport infrastructure in the future to make it easier for autonomous vehicles and safer for passengers. Traffic lights could signal the car by radio to stop and construction site operators are legally obliged to send lane change geometry to the vehicles.
And even if there is a problem despite standardized roads: autonomous vehicles will be operated by companies that can in case of doubt intervene remotely. The car basically operates defensive - if it doesn't know what to do, it phones home and a human takes over.
When launched autonomous vehicles will be in the minority. In unclear situations it makes sense to follow the drivers' instructions, e. g. if road markings are no longer visible due to weather conditions. Experience has shown that the drivers themselves know which route can be driven accident-free!
Learning in virtuality
Nevertheless, remapping of the world is important and will even enable autonomous driving without a map after all. The continuously better imaging of the world with many different sensors and more precise temporal variability modelling in these maps is permanently progressing. The ability to store and process huge amounts of data will lead to an ever-increasing degree of realism in mapping.
New generations of autonomous vehicles will learn how to drive in realistic, virtual cities - including traffic lights, construction sites and unforeseen events on our roads. The human passenger then may read about completely different AI topics, while an Artificial Intelligence matured in many years of simulation finds its way to work - even without a map.