Driverless cars of 2100
There’s inevitably a lot of hype around what will be the transport of tomorrow. Most likely, you’ll be able to request “Uber”-like public transport services, like multi user driverless pods; you’ll hail them electronically and the nearest one that’s heading your way will come by, pick you up and take you where you want to go. The transport system should, if this happens, work much more efficiently and hopefully traffic jams are going to be a thing of the past. Vy Nguyen has been to see the transport of the future for herself…
Ben - Hi, I’m Ben Peters. I’m co-founder and BP product at FiveAI. We’re at our proving ground so the facility where we build and test our vehicles. This is a standard Ford vehicle platform, so it’s the Fusion platform which is a light electric hybrid that we’ve modified with a bunch of autonomous kits to make it drive itself. If I show you around the vehicle…
Vy - The car of tomorrow looks a lot like the car of today but with some rather funky add-ons. The roof is decked out with various cameras and sensors, and sci-fi looking antennas align the bonnet and boot, while the interior is largely full of flashing computer power. And these adjustments help the car to see and think…
Ben - One of the first steps we perform is the localisation step and localise ourselves to a Primap. We use a bunch of different sensors to do that so we’re using the lidar and the vision sensors to do that, and we do also occasionally use GP, which is partly what you can see on the car - the antennas on the rear. Once you’ve localised yourself one of the next things you need to perform really is recognising everything that isn’t on the map so what we refer to as dynamic objects. So that is all of the pedestrians, all of the cyclists, all of the vehicles. And we need to be able to classify them as that, so accurately determine that they are indeed pedestrians, or cyclists, or vehicles, and even what type of vehicle they might be. Are they a lorry or are they a sports car, etc? And to position them in 3D space and to have an idea of what they’re pose is and what they’re velocity is.
Then we need to actually determine what’s likely to happen next in a scene. By that I man what is the likely action and interaction of the dynamic agents in a scene. To do that with any confidence we need to have some understanding of how those dynamic agents tend to behave. We do that by learning typical behaviours. We learn from the data that we capture with our vehicles and we learn from the data that we capture from CCTV footage. We learn how actors tend to behave and how they tend to interact with each other. And then in run time in our vehicles, based on the learning we have on how these agents tend to behave, we play Ford’s multiple potential futures and try to analyse a path through those potential futures that leads to us safely getting to where we want to go.
Vy - But even with thousands of hours of data and machine learning, we’re still not close to matching the prowess of the human brain…
Ben - If we look at some of the science problems we have to solve, the classification performance of things like cyclists at the moment, best in class science is something like 75 percent precisions, meaning we’re still missing about a quarter of cyclists on a frame by frame basis. Novel science is need to get classification performance of many of the things that we care to identify to get to the level that it needs to be to be safe. And then even on the predictions and path planning side really being able to accurately predict how agents interact, and to do that safely over a reasonable time horizon is still and unsolved problem.
Vy - Well the science has a ways to go before we can unleash these cars on our roads. The proving ground provides a safe place to test and troubleshoot. So, of course, I had to have a ride in one of them…
Jamie - Hi, I’m Jamie Lowrie. I’m one of the development engineers here at FiveAI.
I’m currently parked at the start line. I’m just going to press the engage for the autonomous mode so we’ll see the system take control. I’m completely hands free at the moment. We’re coming up the first corner, which is a switchback right then left over a brow, and it’s just controlled itself over the top of the hill. We’re just about to come to a stop here so you’ll feel the brakes come on as we come to our end way point and I’ll disengage… that was the system disengaging.
Vy - Being driven around without a driver was surreal. But what’s the plan for when these cars hit the roads alongside you and me? Back to Ben…
Ben - What we’re aiming to do is to build this autonomous technology into our service vehicles and delivery urban transportation services that are more attractive than driving your own personal vehicle, and can be delivered at a significantly lower cost. These will be shared services so we think if we can get the service design right, it will encourage people to give up their personal cars to share these vehicles and, therefore, reduce congestion and the environmental impact of congestion.
Most people don’t actually enjoy commuting in their own personal vehicles. They tend to be stuck in traffic jams and they buy these fairly expensive items - the most expensive people buy after a house and have them sat depreciating on their driveways for 94/95 percent of the time. We burn up the world’s resources in creating these cars and then just have them sat rotting in car parks. So that is kind of both the economically and the environmentally low hanging fruit that we want to replace with our consumer service.
Vy - Maybe in the year 2100 human driving will be completely obsolete. With the far more energy and time efficient driverless cars making traffic jams and accidents a distance memory. But before then, driverless cars need to share the roads with driver full cars which presents it’s own challenges…
Ben - For many many years, the autonomous vehicles that we develop will be sharing the roads with human drivers, and human drivers have a certain expectation for how other human drivers behave, and so we need to be cognisant of that. If we start to introduce behaviours that look very different to how human drivers behave that could cause a safety problem in itself. We need to drive in a way that is predictable for the human drivers that we share roads with.
That said, human drivers often take risks which they shouldn’t take. A classic example would be human drivers take blind curves faster than they should. Those are the types of behaviours that we won’t take with our vehicles and so our vehicles are constantly, several time a second, reviewing the risk in a scene, reviewing the confidence that we have in a scene, and reviewing the visibility that we have and modifying our speed accordingly and that’s not always something you see with typical human driving.
Chris - Not something that I’d ever considered actually that the driverless car will drive too well for human drivers to anticipate what it might do. That’s ironic isn’t it.
That was Ben Peters and Jamie Lowrie, from FiveAI, talking to Vy Nguyen taking a ride in one of their cars.
Mike, Pitts who is from Innovate UK is still with us. Mike, what’s your vision for the transport services of 2100 and also especially long distance transport?
Mike - I think within cities, as you heard in the piece there, there’s much more optimised ways of moving around so we’re using fewer vehicles to move people much more efficiently. What’s more interesting is the intercity kind of connections, we’re very excited about technologies like hyperloop probably coming along by 2100 and that’s essentially the kinds of things you’ve seen in movies. Big pods flying down tunnels that have had the air removed from them so that there’s no air resistance and we can get from city to city in minutes.
Chris - So we’ll have even more time to spend on admin wouldn’t you say Mike?
Mike - Or in the pub!