Self-driving buses might come sooner than self-driving cars
Automotive automation is not easy
Self-driving cars are an enormous technical challenge.
To help quantify the challenge, the Society of Automotive Engineers has created a series of autonomy levels
Level 1 involves automating a single element. Level 2 involves automating several elements; like lane control or parking. Level 3 involves the car taking control of safety-critical functions. Level 4 involves full autonomy in a defined geographical area while Level 5 is the utopia of a car able to go anywhere, anytime.
Yet despite extraordinary sums invested in autonomous vehicles, progress has stalled since Tesla introduced its “Autopilot” feature in 2015. Autopilot is level 2 autonomy; the car can do certain tasks by itself but with a human at the wheel. At the time, Elon Musk promised full autonomy in 2017. He restated the 2017 target in 2016. In January 2017, he said on Twitter that full self-driving capability would be available in “3 months maybe, 6 months definitely”.
6 months, 4 years ago
Four years on, we’re still waiting. Yet earlier this year, Musk confirmed (again) that full Level 5 autonomy would be available on Tesla by the end of 2021.
Given that Tesla hasn’t progressed beyond Level 2 since 2015, this seems about as plausible as most Elon Musk statements.
Yet you can see why Musk would be confident. The basic technology of self-driving has been working well for many years. In Phoenix, Google sister company Waymo is offering Waymo One, a fully autonomous ride-hailing app. Here’s their promotional video, explaining the technology behind self-driving cars and showing what it looks like to ride in a car with no driver:
It’s all very compelling, but it also highlights why fully autonomous cars are so hard to make work. As the video says, the system is based on the car’s computer making predictions on what other road-users are going to do. Those predictions are built up from the experiences the system has: drawn from millions of miles of travelling.
But the problem is the edge cases that humans can process easily and computers struggle with.
As an example, let’s imagine that a self-driving car finds itself presented by a group of three children carrying home a school project. The project is large, made of cardboard and has got wheels on it. The kids walk inside it to carry it and it’s so big it needs one of them to be walking in the road. A human driver would immediately recognise that it’s kids carrying cardboard; but a self-driving car has never seen anything like it before.
Indeed, the staggeringly ambitious necessity of Level 5 self-driving cars is that computer vision must be able to recognise anything, anywhere, in any environment and in real-time; and the car must make an instant and accurate prediction of what will happen next. With lives resting on the accuracy of the prediction.
As soon as you process the enormity of the task, you realise why progress towards autonomy has stalled.
The Waymo One fully automated cab service is impressive but it still struggles. Here’s a great video of a Waymo One cab being foxed by traffic cones that are more spaced out than it seemed to expect (it’s a long video; the interesting stuff starts from minute 12):
Obviously, Waymo One is still learning and overall it’s hugely impressive. But it’s also limited to the East Valley of Phoenix, Arizona; a section of an American city famous for its straight, wide roads. This is a very sensible location to pilot this project - but it’s a million miles from me being able to get into a self-driving car outside my house in Walthamstow and head to the in-laws in Halewood.
Move over cars; let the buses handle this
Lost in the billions invested in cars is the fact that buses may be a far more suitable use-case for this technology.
Unlike cars, buses aren’t meant to be able to go anywhere. So there’s no need to try to achieve Level 5 at all. A bus is expected to stick to a predictable series of roads.
And whereas having to employ remote operators to assist cars that get into trouble would make driving more expensive, replacing drivers on vehicles with operators in offices would make driving buses cheaper. Autonomous buses could be driverless, but if a bus ever got bamboozled by a cardboard kids’ school project, someone could simply log in and drive it remotely from a control centre until it was happy again. Yes, that is also true of cars, but the difference is that there are 31 million cars and around 90,000 buses. Having every self-driving bus allocated a control-room operator is practical in a way that isn’t true for self-driving cars.
Now, I’m obviously not suggesting that a peak-time double-decker would run without staff on board. But busy buses with lots of boarding and alighting already cover the cost of their staff. Quieter buses need subsidy; and are more practical to automate.
For example, the bus that serves my street in East London, the W12. The W12 wiggles and winds its way through Walthamstow to Wanstead (I really enjoyed writing all those Ws…). Much of the route is hail and ride. It uses small buses and, even before Covid, was always lightly loaded. In December 2017 the frequency was reduced from 20 minutes to 30 minutes. Given that it’s hail and ride, there are no bus stops, so that often means long, slightly tedious waits by the side of the road.
The W12, captured by Google, as it turns into my street
But an autonomous W12 could go back to every twenty minutes. Or, indeed, every 10 minutes? After all, the cost of energy and depreciation are tiny compared to the cost of the driver. Now, you might say that the driver has a role beyond driving. But the W12 is not community transport. The driver doesn’t assist passengers on and off, and doesn’t collect fares. They are isolated in their perspex box today and it’s hard to see the travelling experience on the W12 being all that different if they weren’t there.
Now I can hear all the concerns. There are concerns about boarding and alighting safety, there are concerns about anti-social behaviour, there are concerns about loneliness, there are concerns about revenue protection, there are concerns about customer acceptance, and probably many other concerns I haven’t listed here. And these are in addition to the obvious concerns about the ability of the computer to actually drive the bus.
And there are, in places, obvious contradictions: the economic benefit of self-driving buses would be greatest on relatively lightly loaded night buses for which it is hard to recruit drivers but these are the very buses on which the drivers provide the most visible reassurance.
But there may be solutions to these problems. Anti-social behaviour fears are probably overdone; there were once worries about taking the guards out of tube carriages but few people suggest bringing them back. Revenue protection may be dealt with by computer vision. Loneliness concerns are real but the levels of engagement between driver and passenger on most bus routes is probably not a significant alleviation at the moment; certainly not compared with the improved connectivity and social interaction that expanded bus networks might facilitate. Boarding and alighting on quiet bus routes feels like something that computer vision will be able to manage though I may be wrong.
Certainly it doesn’t feel so certain that these problems cannot be solved that automating bus services should not be a major focus for the bus industry.
Unlike the endless discussions about driverless tube trains (which are largely irrelevant: it is inconceivable a tube will ever enter a tunnel without staff onboard; so driverless tubes are virtually identical to the already automated trains on four tube lines), driverless buses have the potential to transform the economics of the bus industry and facilitate a big expansion in the service.
(And don’t worry about employment: it’s not about sacking drivers. It’s about enabling more buses. Amazon operates automated fulfilment centres but needs 1.3 million staff to do so, because automated fulfilment has so expanded the market for home delivery. It’s the same idea for buses. The expanded bus networks that would previously have been unprofitable will still require plenty of staff on the busy, crowded routes - which, if we get this right, will be an increasing proportion of the total as the network effect flywheel spins)
In that context, it’s somewhat surprising that the potential for automation gets no mention at all in the National Bus Strategy. Given that the strategy found room to describe how bus maps should be drawn, whether or not bus routes should have letters in the numbers (they shouldn’t, FYI) and whether or not flat fares are better than variable fares (they are, we’re told), it’s a bit odd that there isn’t a single mention of the biggest change taking place in transportation since the invention of the car. Especially in a document billed as a ‘strategy’, you’d expect some consideration of medium and long term technological change.
But just because the bus strategy forgets about autonomous vehicles, doesn’t mean that the technology has ceased to exist.
And, in fact, autonomous buses could be the key that unlocks the paradox of the bus strategy.
The key feature of the bus strategy is that it wants bus services that are so good they’ll make you weak-at-the-knees but with virtually no money to pay for them, as it assumes they’ll be able to run commercially. Networks that are comprised of a mixture of manual (on busy, crowded routes) and autonomous (on quieter routes or at quieter times) could be the solution to this very conundrum. The problem is the breathless timescale of the bus strategy.
Join in
Britain’s first autonomous bus service, which ended its pilot on Tuesday, was a collaboration between Aurrigo (a subsidiary of automotive engineering group RDM) and the Greater Cambridge Partnership. Innovate UK stumped up the cash. The traditional bus industry was nowhere to be seen.
(The buses were also of such stupendous, overwhelming ugliness that it’s no wonder the established industry didn’t want to get involved. As well as the prize for Britain’s first autonomous bus. the Cambridge trial should definitely win the prize for Britain’s ugliest).
Gosh, that’s ugly
But the reality is that urban buses are a perfect use-case for autonomous vehicle technology. They stay in one area. They use predictable routes. They move slowly. And whereas adding human resources for remote monitoring of cars harms the economics of motoring, replacing a driver per vehicle with a monitor per depot is of huge benefit to the economics of buses.
In that context, now is the time for those institutions with a remit to think long term (the DfT and the PLCs) to engage.
I very much doubt that Tesla will achieve level 5 automation this year. But there is no reason why we couldn’t be piloting our first autonomous bus routes this year, and holding open the door to a transformation of the economics of the bus sector.
