How Cambridge company FiveAI will change commuting forever with its autonomous service vehicles
Co-founder Ben Peters says fast-growing AI firm is already testing its driverless vehicles on the streets
In a decade, commuting by driverless vehicles will be commonplace.
In two decades, driving your own car will be rarer than taking an autonomous ride.
Those are the predictions of Ben Peters, co-founder of Cambridge company FiveAI, which is among those leading the way in this field.
“We are building autonomous vehicle technology that we will deploy in service vehicles and deliver as a transportation service direct to consumers,” Ben tells the Cambridge Independent.
“Fast forward to 2022 and people will be able to tap a button to get a ride and a shared autonomous vehicle will turn up, pick them up and take them where they want to go.”
FiveAI was launched just three years ago, but already it is road-testing its autonomous vehicle technology on the streets of Bedfordshire, as well as on a track circuit built for Vauxhall.
“The idea is to compete with the Googles of this world,” he says. “The interesting segment is being able to deliver this in an urban environment and replace commuter cars. People are driving to work and typically only one seat is occupied.
“The opportunity is delivering a service that is so attractive that commuters give up their cars.
“They’ll get a point-to-point journey that is delivered at a cheaper price than a bus ride.”
FiveAI aims to deploy eight-person electric minibus vehicles on the streets of a south London borough in four years, ushering in a new era of transportation that, Ben believes, will be utterly transformational.
“It’s economically interesting but the socio-economic and environmental benefits could be vast as well,” says Ben, who was part of the panel for Mills & Reeve’s Question Cambridge event.
“There are 40,000 people each year having early deaths as a result of air pollution.
“All the vehicles we’ll operate will be electric and because they are autonomously controlled, we can depot them in the same place, which means we can adopt electric vehicles much quicker than if they were personal vehicles, where you need charging infrastructure all over the place.”
The increase in productivity will also be considerable.
“When I commute, what I really want to be doing is reading my emails, or a book, or talking to my daughter on the phone. There’s no pleasure in the driving because you’re stuck in traffic.
“We are typically wasting 230 hours a year behind the wheel. If we can get this right and free up that time for people and do it in such a way that creates an interesting business and introduces some environmental benefits, that is a utopian combination that will see a huge take-up over a short period of time,” he predicts.
Under FiveAI’s service, passengers would order and pay for a vehicle via their phone and it will pick them up close to home, place of business or railway station. But Ben acknowledges the service also needs to work for the elderly and less able, who may not wish to use a phone to pay,
“People may be able to hop on and off and use their Oyster card, or perhaps they pre-book over the phone,” he suggests.
When the Cambridge Independent catches up with Ben, he is at the Labour party conference. The company is also attending the Conservative party conference.
“It’s important for us because we do need some regulatory changes to introduce our services. We need a change to the Road Traffic Act, because at the moment you need a driver to be in control of a vehicle at all times. It is slightly ambiguous what could be called a ‘driver’ but we don’t want to live in a world with that ambiguity,” he says.
For its road testing, there is a safety driver behind the wheel.
“We’re demonstrating a prototypical route between our test track through local roads to a train station,” says Ben. “Although that’s close to our test track near Bedford, it’s a good example of what we’d expect to see in the outer boroughs of London. It has traffic lights, roundabouts and pedestrian crossings.”
The road testing has revealed a host of interesting challenges.
“You can’t sit at your desk and write a specification for an autonomous vehicle,” says Ben, who is VP product for FiveAI.
“It’s very hard to specify the dynamic range for the cameras. It’s very hard to specify the types of behaviours you’ll encounter.
“We have a formal ontology that describes the world. We have models that described effects like rain, snow and sunlight and we have behavioural models that describe the behaviour of cyclists and cars and buses.
“We run those models in a loop while our cars are out driving.
“Then, when something in those models is unrepresentative of what we observe in the real world, we update those models and design solutions against them,” Ben explains.
While there will be understandable concern over safety, Ben points out that our roads are not safe now.
“We seem to have built up a tolerance that people lose attention when they are driving and people get hurt,” says Ben. “When humans are paying attention, they are very good drivers and it’s a trivial task. The problem is, we’re not very good at paying attention.
“No-one has found a very good solution to that. It’s our natural desire not to want to pay attention to trivial tasks.
“An autonomous vehicle is nothing like general intelligence. It is like a pigeon brain to a human mind. But because they can pay attention, they can be much safer.”
Danger could come from another source, however: hackers.
“It’s something that has to be addressed. We do need to be able to take remote control of these vehicles.
“For certain edge-case scenarios where the vehicle gets confused by something in the road and doesn’t know whether it can proceed or not, we need to have remote control. And that potentially opens up a vulnerability.
“But these problems are tractable. In the context of all the problems we’re solving, it’s not the most difficult.”
The company is exploring 30 miles of routes across Croydon and Bromley for its launch route.
“We’re in the mode of understanding their complexity. We’ll deploy on routes that are to some extent constrained. We’re not going to drive Oxford Street on day one.
“Certain problems are harder than others. Traffic lights are easier than junctions without them. All these things we’ll need to deal with in time but they are not for the launch routes.
“Assuming we can find routes that are economically interesting and sufficiently constrained, we’ll probably launch routes in 2022,” says Ben. “Cambridge is definitely a city we’ll come to at some point but the economics of it mean we need to raise a lot of money, so we need our first market to be one that serves a large market so that we can rapidly expand.”
It is a daunting challenge, but FiveAI’s founders are not ones to shy away from that.
“As a team, we all came from deep tech projects in the past and have done several start-ups together and sold them successfully,” says Ben.
CEO Stan Boland led Icera, Element 14, Neul and Acorn and sold the first three for more than $1billion.
“Our mission with this one is to build it to a point where it is self-sustaining and we don’t have to sell it. We all feel we’d like to build a company that stays British for the long-term,” says Ben. “We’ve raised about £21million over a couple of venture capital rounds to date. We also got an $11million grant from the UK government to trial our technology in London. We’ve grown the team to about 125 people.
“But we’re really just getting started. It’s a huge mission we’re taking on. By the end of 2020, the team will be circa 400 people.”
Major investment will be required to make this vision a reality. But if they get it right, the prize is enormous.
“There is $1.2trillion spent on personal mobility in Europe every year and about 70 per cent of that is in cities,” says Ben. “If we can solve these problems, the market is absolutely huge.
“At the tail end of the 2020s, it will be perfectly normal for people to commute in autonomous vehicles. But it will be a heavily mixed traffic environment, with many people driving their own vehicles.
“By 2040, people driving their own vehicles will be rarer I think than taking autonomous vehicles.”
What’s a driveable surface? How do you spot a child in the road? And how do we prove it’s safe? Three technical challenges to making driverless cars a reality
FiveAI’s technology needs to solve three broad challenges to enable autonomous vehicles to operate on the streets.
“The first is perception,” says Ben. “This is a series of computer vision technologies that we need to develop, such as going from 2D images to 3D to detect depth – which is called visual geometry – then, from a semantic perspective, having segmentation of the scene, so we understand what is the drivable surface, what is our lane and what is the opposite lane, and what are the cars, cyclists, buses.
“There are techniques that have made a lot of progress in computer vision competitions over the last seven or eight years that rely on deep neural networks to solve those problems. Those are definitely helpful, but each has their own failure mode. They tend to be quite fragile.
“A lot of the problems we’re tackling are about making this a lot more robust and accurate by improving the precision recall performance.
“There’s a world of difference between interesting demonstrators that can identify cats in images versus the necessity of accurately detecting when there is a child in the road and stopping.”
The second major area is in predicting what is about to happen – very challenging in the urban setting on which FiveAI is focused.
“On a highway, everyone is moving in pretty much the same direction at high speed and the degree of freedom of motion is quite limited,” says Ben. “In an urban environment people can do just about anything. People can jaywalk; cars can do U-turns, so being able to accurately predict how scenes unfold over a three, five or 10-second horizon is key to be able to plan safely through a city. Understanding when you can’t predict, and therefore the best thing to do is slow down, is one of the key areas we are working on.”
The third major area of focus is complex system verification.
“A lot of the techniques we use are black box techniques,” says Ben. “For a lot of deep neural network techniques used for perception it’s very difficult to interpret how they are making their classification decisions. If we don’t understand how these components reach their decisions, it’s quite difficult to understand the failure modes.
“We have to build systems that are safe but we have to prove that they are safe, which is a difficult problem that spans novel science and verification techniques that we’re used to finding in the semiconductor space.
“One of the problems is pulling research that is just about proving fruitful in university into rigorously engineered solutions that we can safely put on the road.”
A key job for the co-founders has been recruiting world-leading experts.
“In the first year of the company, I must have read 1,500 academic papers. It’s super interesting but a massive learning curve – just to be able to understand the science sufficiently to hire people who are experts in it!” says Ben.who previously worked at Internet of Things business Neul and NXP Automotive.
“In the first year of the company, I must have read 1,500 academic papers. It’s super interesting but a massive learning curve – just to be able to understand the science sufficiently to hire people who are experts in it!”
Among the team is Dr Ram Ramamoorthy, VP prediction and motion planning, a research leader at the University of Edinburgh with a PhD in robotics, and computer scientist John Lusty, VP simulation, who lead the virtual reality and AAA games development at Facebook/Oculus, Square Enix and Ninja Theory.
How Cambridge could become first UK city to have autonomous shuttle service
Cambridge could become the first UK city to have an autonomous shuttle service as part of its public transport network.
In February, £3.2milllion of Government funding was secured to develop trial vehicles by Smart Cambridge, which is led by Cambridgeshire County Council, and the Greater Cambridge Partnership.
Coventry-based Aurrigo, the autonomous vehicle division of RDM Group, will develop the vehicles and last year trialled an autonomous four-seater ‘pod’ in Cambridge.
The money will fund the building and testing of six 10 to 15-seater self-driving shuttles that will operate on the southern section of the existing guided busway. Initially this will be an out-of-hours service, when ordinary buses are not running.
As the busway is segregated from general traffic, but runs near residential and employment centres, it offers the perfect testbed. If successful, an autonomous shuttle service could run in the early mornings, late evenings and at weekends, filling a void for shift workers, the night economy and weekend shoppers.
An initial service is envisaged between Trumpington Park & Ride and Cambridge Railway Station, via Cambridge Biomedical Campus. Self-driving vehicles could then be rolled out further to link science and business campuses to each other and rural travel hubs.
The project is due to get under way by the end of the year, with prototype vehicles tested in late summer 2019. The first passengers could use the service in summer 2020.