Whilst driverless cars are advancing rapidly they’ve got a long way to go before they can understand and embrace the unwritten rules of human driving behaviour. David Fowler looks at how the LAMBDA-V project is developing the technology to help this happen.
Last month the government announced moves to update regulations for testing autonomous vehicles, aiming to pave the way to allow fully self-driving vehicles to be introduced on UK roads by 2021.
Observers believe that cars that will be able to steer, brake and accelerate by themselves on sections of motorway could become a reality within a few years.
But what about driving on the residential streets near your home: when will an autonomous vehicle (AV) be able to manage that without human intervention? How often do you find yourself in a situation where the rules governing driving – a combination of the law and the Highway Code – don’t really apply, and you have to use your own judgement? What will AVs do in such situations?
A new project, LAMBDA-V (Learning through AMBient Driving styles for Autonomous Vehicles), is looking into that question. Starting last November, a one-year £244,000 feasibility study, with funding from Innovate UK, is collecting data from “ambient” driver behaviour to see whether it can be codified into a set of rules for AVs to follow.
The project is led by machine learning specialist CloudMade, with telematics and Big Data analytics firm Trakm8, traffic modelling specialist Aimsun (a Siemens company), and Birmingham City Council, which wants to know what the implications will be for how its road network is operated. Andy Graham of White Willow Consulting is project manager.
An observed drive Graham took from his home to the M25 illustrates why the project is needed. In the 3.1 miles, there were 23 situations in which he took decisions that could prove challenging to an AV because they were not strictly in compliance with the Highway Code.
Incidents included a car parked just before a road junction, making it necessary to approach on the junction on the right hand side of the road; a place where another parked car made it necessary to drive with two wheels on the kerb to get past; a stretch of road wide enough for only one vehicle, with a blind bend and just one passing space midway along, posing the dilemma of what happens if you meet something coming the other way; and a junction where, if you want to turn left and then right at the junction immediately following, it is necessary to make the left turn in the right hand lane in order to be in position to take the right turn.
In situations such as the narrow road, human drivers will signal by hand or by flashing their headlights to invite another vehicle to go first. It’s unclear how AVs will react to such situations and driver behaviour.
With the average age of the car fleet around eight years, it would take at least that long before half the fleet becomes autonomous even if all new vehicles were AVs. So AVs will have to mix with humans for a considerable period. “Humans drive vehicles: there is a need to understand human behaviour and how AVs react, to tailor the early AVs so they drive like humans,” said Graham. If they don’t, it could be confusing and unsettling both for human drivers and for AV passengers.
The partners in LAMBDA-V have complementary skills to bring to bear on the problem.
Harsha Bagur, CloudMade head of data protection and solutions architect said: “At CloudMade, we use machine learning to learn human behaviour patterns and use them to predict driver behaviour.”
The company learns the driver’s behaviour based on observation of their actions and by aggregating relevant data from the vehicle. To understand the context of specific driver behaviour, it is important to understand external factors such as the weather, traffic and road conditions. CloudMade’s technology considers multiple factors and data points to learn and predict the human driver’s behaviour in different conditions. For example, said Bagur, the car might learn that if a journey starts at 8am the driver is likely to be going to work, whereas if it starts at, say, 7.30 with an additional passenger at the rear, the destination is likely to be school. It will learn the routes to frequently visited destinations.
Lambda-V will allow CloudMade to study a greatly enlarged pool of data, and also build capabilities in building rulesets for autonomous vehicles. “What we’re trying to do here is to observe how humans drive – collect data on this behaviour and see if we can build appropriate rules and models for autonomous vehicles. The key is also to model this behaviour at critical junctions and road segments that are contentious for autonomous vehicles to take decisions. Simultaneously, it is important that the road networks and the legal frameworks evolve to allow the early adoption of autonomous vehicles,” said Bagur.
Trakm8 produces a plug-in telematics device which fits into a car’s OBD2 on-board diagnostic port. This “dongle” contains an accelerometer, GPS and a mobile phone SIM, and measures driver behaviour as demonstrated through acceleration, braking, and cornering, as well as recording data from the vehicle’s Canbus, such as whether the windscreen wipers are in use or when the ABS comes on. The telematics data is already collected and used in sectors such as young driver and use-based insurance, fleet management, and by the AA’s roadside recovery teams for locating vehicles which have broken down. Trakm8 has around a quarter of a million devices reporting to its servers.
Aimsun develops mobility modelling software that integrates all modelling levels, from wide-area macroscopic modelling right down to the microscopic simulations showing the behaviour of each individual vehicle.
Birmingham City Council is interested in how the road network might need to be adapted for AVs and what policy changes might be needed. For example, will parts of the network be effectively impassable to AVs because they are too narrow or lined by parked cars, or both? What will the implications be if, say, a third of the fleet is AVs and they all have to divert to the main road through a town centre instead?
Graham summed up the collaboration: “Trakm8 collects data, CloudMade uses it to build rules, Aimsun uses the data for modelling and Birmingham is interested in the outcomes.”
There are vehicles that can drive on motorways now. Driving down residential streets is more difficult. If the AV can’t get to your house and park in your drive, that will be a poor user experience
Trakm8’s participation potentially enables a great deal of data to be collected, which has to be anonymised before it can be analysed, in line with the EU General Data Protection Regulations (GDPR). At present, data is planned to be collected from 1,000 vehicles. CloudMade will process this to see if it can be distilled into rules for AVs to follow.
The task is complicated by the fact that different parts of the country have different conventions. In London drivers are reluctant to let joining traffic out of side roads; in Devon they are more likely to, and in Jersey drivers take turns.
Aimsun will be able to process the rules that emerge, and model the changes in traffic patterns if AVs were to behave differently from cars driven by humans, particularly if they were to drive more cautiously. “We’re also exploring to what extent data connection between vehicles would help, for example if another vehicle is coming around a blind corner and can let you know,” said Graham. “In a mixed fleet, would you have to get to the point where everyone’s connected?” But again there is the question that even with connected vehicles, a human driver might react differently to the information than would an AV.
To allow study in extra depth, data is also being collected from the A45 between the junction with the M42 near the National Exhibition Centre, and Birmingham city centre. This section of road was used for the GLOSA (Green Light Optimised Speed Advisory) adaptive traffic light trial, which advised drivers what speed to travel at to pass through the next green light, so it is already equipped with CCTV and has a number of participating vehicles providing data. This will allow correlation when unusual driver behaviour is recorded – such as many drivers swerving at the same point (suggesting a pothole exists), travelling more slowly than expected, or more than the usual number turning off the route. It will be possible to look at CCTV footage for the corresponding time to find out the cause. “We’re mining data that would otherwise be thrown away,” Graham said.
The study hopes to answer a number of questions. Can CloudMade devise rules for its car manufacturer customers that will make AVs behave more like human-driven cars? Can Aimsun use those roles to model behaviour of AVs better? Can Trakm8 use the data to further improve both its telematics hardware and driver behaviour algorithms? If these questions can be answered successfully, highway authorities such as Birmingham City Council will gain a better understanding of how AVs will affect the use of their networks. And CloudMade, Trakm8 and Aimsun will be able to develop innovative products that the UK can export around the world.
Finding satisfactory answers could also be crucial to the acceptance of autonomous vehicles. Will early adopters be so keen to spend £80,000 to acquire an AV if it can’t take them all the way to the drive of their house, or can’t reach their child’s school because the road is too narrow? What about the use case of guaranteeing mobility for the elderly if AVs are unable to access hospitals or care homes, or the prospect of using AVs for deliveries if every address cannot be reached?
“The early days of AVs will be the real test ground,” said Graham. “There are vehicles that can drive on motorways now. Driving down residential streets is more difficult. If the AV can’t get to your house and park in your drive, that will be a poor user experience.”