Home / IT & Computer Science / AI & Robotics / Building a Future with Robots / Learning from good drivers
Learning from good drivers
Sandor presents his research into driver experience based decision-making autonomous cars (DEBDAC).
Autonomous driving is bound to come in the future. However, when it comes, it may not be entirely safe. People talk about the prospects of autonomous vehicles saving the 90% of crashes that are human-led. However, human understanding leads to that machine understanding is not perfect either. What we need to do is, we need to make sure that the understanding and fit of these vehicles into the road architecture and their movements doesn’t cause a shock to the system and introduce more chaos than it’s actually trying to save. So what we need to do is we need to make sure that there is a firm understanding of risk.
Decisions are made based on the most available and the current information as best as possible, so that autonomous vehicles can fit in reliably into the roads of today. I’ve been looking for a number of years now at human driving behaviour. We gather many hundreds of thousands of vehicles in real time, every time they’re moving in subsecond detail, of how they’re moving in terms of the GPS, accelerometer, and a range of other sensors depending upon the device fitted into the vehicle. This gives us a lot of information about how people behave, down to individual metre segments of every road in the world. However, it does lead to an understanding of the individual vehicle, not actually the vehicle in context.
There is still a lack of certain information that we can’t tell if a vehicle is having strange accelerating and braking behaviour because of vehicles in front of them, or actually, it’s just the driver driving a little bit erratically, because we don’t have the context of wider sensors. Autonomous vehicles bridges that gap and allows us to examine more sensors and more context around how people drive to get a better picture of what is safe, how to make better decisions. And using this information in combination with traditional systems for making decisions, such as vision-based systems - is there a car in front of me? Therefore, slow down.
You can also start to take into account other information, like actually, do all other vehicles slow down at this point? Is this normalised behaviour, or am I acting outside the norm of what human drivers would do? These can lead to better decisions made inside autonomous vehicles. We gather this information on behalf of insurers with their opt-in customers who choose to opt into policies that allow them to get significant savings on their insurance in return for giving data about how they drive.
It eliminates things like whiplash fraud, and so on, thus allowing insurance costs to be lowered, which is a saving for the insurer, a saving for the driver, but also introduces technology that’s gathering essential data about the condition of roads and the behaviour of people as they move. We’re looking for behaviour that is predictive of risk ultimately, which in a human driving capacity is the drivers that drive more aggressively, the drivers that drive more irresponsibly. They drive outside the normal envelope of most other drivers. And those drivers, when they have certain traits of behaviour, they can correlate to claims from other drivers that have similar traits and performances. And from that, you can use that to help educate.
So you can actually feed back and help to improve drivers. You can actually inform them to become more responsible. You can actually improve the general lot of an insurer’s book of policyholders. They can actually become better. The information is also valuable in many other areas. It can help with a whole range of things, from where to place the next IKEA, the next store, the next thing to inform about road structures. Is the signalling on that lamp post timed right? Or is there certain other things that are operated and optimised as best as possible about the road network?
What we’ve tried to do in this project that we’re starting now is, we’ve tried to look at the data to see how we can use it to help improve autonomous driving. For some time, we have been studying autonomous driving at the university, and we have come to the recognition that it’s a very hard problem as far as decision-making is concerned. The realisation was that this can’t be done just by LIDAR and computer vision where the empty spaces are. There is more complex decision-making at hand. And then we discovered The Floow Ltd are gathering data as vehicles go on the road. We realised that there is an opportunity here.
And the opportunity is to extend perhaps the data gathering, what Floow Ltd is doing currently, and look at instantaneous data about a scene - what is happening with the other participants of the traffic - and derive from that the kind of decision-making that human drivers do. The DEBDAC is a driver experience based learning system for autonomous cars. The title says what it is. We are gathering data about drivers’ experiences. It drives hundreds of miles on various roads, not only on motorways, on small roads, in urban traffic, in dense traffic, in traffic jams, all kind of stuff.
And we were going to measure what other vehicles are doing, what the pedestrians are doing, what is the weather, what traffic signs, what roadblocks in terms of road repair, what potholes appear, actually, on the road surface and so on. And we gathered this data into a huge database, and we analysed the decision-making of the human driver. And then we wanted this system to be transferred, basically to the autonomous car decision-making. The purpose is to make decision-making in autonomous cars better. So at the moment, Google has their autonomous cars on the road, and also various universities have various autonomous cars on trial.
And we have reports about some kinds of consequential accidents which are not necessarily because the autonomous cars make some mistake, but it doesn’t understand the human context. Most often, the problem is that people can’t stop suddenly when the autonomous car stops. And also, when the car should be polite, it may not do exactly the right thing. It’s too abrupt. So it doesn’t have the social situational awareness. We need to monitor that. Insurance companies are definitely very interested in that, that we have a very smooth transition to autonomous car usage.
A car manufacturer should be very much interested in autonomous driving capability decision-making, which in effect adapts itself to the human condition, so driving when lots of other human drivers will be around still. And this is a key factor to reduce the number of accidents. In this context, I strongly believe that autonomous cars, in the future, will drastically reduce the number of accidents on the road. And the public will quickly recognise their benefits in terms of the accidents. That’s definitely the case. It takes some time for things to mature, but in fact, the sensor technology of an autonomous car can be, in the future, more advanced than a human capable of sensing, and faster.
And that’s quite a promising prospect for the future, that we will have nearly accident-free roads, if it’s possible.
How will autonomous vehicles drive on the roads of today? How can we ensure that they will drive safely alongside humans?
Automated technology in cars could help to prevent accidents and reduce congestion and emissions in cities. However, automated driving is highly complex and requires systems to respond to a wide range of real-world driving situations.
Professor Sandor Veres has been working with telematics company The Floow LTD to develop automated vehicles that can learn from the behaviour of human drivers.
In this video, we’ll first hear from Dr Sam Chapman, Chief Innovation Officer at the Floow, as he explains the types of information that can be gathered from drivers and how this can be used to educate drivers, and improve infrastructure and town planning.
We’ll then hear from Sandor about how this data is helping the University in their current research project, designing systems that use driver experienced based decision-making (DEBDAC).
Reach your personal and professional goals
Unlock access to hundreds of expert online courses and degrees from top universities and educators to gain accredited qualifications and professional CV-building certificates.
Join over 18 million learners to launch, switch or build upon your career, all at your own pace, across a wide range of topic areas.
