What’s the use of a learner taxi driver?

The concept of a driverless “robotaxi” is a powerful vision that will drive the development of autonomous vehicles for years to come. Like any good vision, it’s bold, ambitious and difficult to achieve.

However, autonomous vehicles don’t need to achieve this vision before they deliver significant value. Autonomous driver assistance systems are already saving lives by avoiding collisions when the driver wasn’t paying attention.

But there are a multitude of other opportunities to deliver value to customers and disrupt established industries, long before the day when a robotaxi picks you up from outside a downtown bar.

Self-driving cars provide a vivid example of how understanding and working with the limitations of an AI technology can unlock business opportunities much faster than waiting for the system to mature.

Companies that focus predominantly on technology development rather than having a wider Service Design perspective are likely to miss these interim opportunities. So too, of course, will the companies that don’t fully understand the potential that this AI technology can offer at each stage of its development.

What are the limitations of Tesla Autopilot?

Tesla is widely accepted as the leader in autonomous driving technology, so let’s use their ‘Autopilot’ self-driving technology as the example here. To know what profitable services we will be able to build with autonomous Tesla vehicles, we need to understand the current capabilities and limitations of Autopilot, and how this is likely to evolve. Let’s look at the two different categories of failure, because it’s important to understand the sources and implications of each.

False positives (i.e. crashing)

The most dangerous limitation of Autopilot is that sometimes it will judge an area to be ‘driveable’ when it isn’t, or to put it more simply, it will accidentally drive into things. The main challenge of autonomous driving is that it can be very difficult for the system to correctly identify an obstacle, without producing so many false positives that it would be terribly annoying because it would keep slamming the brakes on for no apparent reason.

Sometimes the difference may be obvious to a human, but difficult for the AI. For example, a large white sign above the highway can be safely ignored. Meanwhile, a large white overturned tractor trailer on the highway might appear quite similar (to both radar and cameras), but it’s actually a dangerous obstacle to be avoided. The fact that Autopilot finds it hard to distinguish is evidenced by the way that it will often brake when approaching overhead signs, but there have also been a couple of serious collisions with overturned tractor trailers when inattentive drivers were using Autopilot.

False negatives (i.e. getting stuck)

Autopilot often fails to understand that it’s fine to drive around a parked vehicle. The mistake is understandable, because driving around a vehicle that’s queueing could be at best impolite, and at worst dangerous. When there’s a driver in the Tesla, this is annoying but can be swiftly remedied by manually driving around. If the vehicle was driverless and got stuck in such situations however, it would quickly create traffic jams and earn Tesla a bad name.

Making the right trade-off

As with designing any AI system, it’s important to carefully weigh the cost of a false positive against the case of a false negative. In the case of autonomous cars, that equation changes significantly depending on whether there is a human at risk, either inside or outside the vehicle. Accidentally wrecking a few empty cars isn’t a big deal, but killing a human is a tragedy.

The optimal trade-off also changes slightly depending whether there is a human inside the vehicle. If an empty car drives cautiously and takes 20% longer to reach its destination, it just means the asset utilisation is a little lower and the drivers behind will be more keen to overtake. If it drives this way with a human inside, they could get frustrated or perhaps even doubt Autopilot’s competence, as it could feel like being driven by a learner driver. I’m not sure I’d choose to get into a taxi that had L-plates on!

How can we overcome these limitations?

Laser beam eyes and invisible fences

Unlike many of its competitors, Tesla has taken a strong stance against geofencing – restricting the vehicles to certain geographic areas – and relying on the high-definition maps you can generate from spinning laser rangefinders (LIDAR). Instead, Tesla is aiming for a general camera-based solution that will work anywhere.

I agree with Tesla that this will be the best solution in the long term. However it’s an ambitious goal and could take a long time until Autopilot can really handle every scenario. Even patient shareholders like me might start to wonder how slow “Elon time” can go. 

Having high-definition maps doesn’t necessarily help much, because the work is constantly changing – it’s a “brittle” solution, a bit like hard-coding. However, restricting operation to areas where Autopilot is known to work well could be a good move, as we’ll explore below.

Teleport a human on demand

There is potential for human operators to remotely control the vehicles, as multiple companies are doing for heavy goods vehicles. However, we wouldn’t want to employ a full time operator for a passenger vehicle, otherwise they might as well be sitting in the car.

So assuming that operators will only “beam in” when the car can’t figure out the situation, the car needs to be able to deal with all emergency situations itself. After all, it’s not realistic to expect a remote operator to react immediately to an emergency situation they have been virtually ‘teleported’ into. Instead the car must be able to slam on the brakes itself, and remote operators beam in only to help it navigate out of confusing situations where it got stuck, such as trying to find the way out of temporary roadworks or navigating safely around accidents.

An automated cry for help

If we start thinking about remote driving, you might think that Autopilot could also ask for help when it anticipates a challenging road situation ahead. However, that’s a lot harder than it sounds. As Andrej Karpathy, Tesla’s Senior Director of Artificial Intelligence puts it, “detecting that a network doesn’t know, and doing it efficiently at test time is still an open problem, in my mind.”

Fundamentally, it’s hard for Autopilot to know when it is going to fail, because autonomous driving is almost entirely a perception problem, not a control problem – i.e. once you understand what’s going on, it’s reasonably straightforward to decide how to steer and what speed to aim for. So if Autopilot ‘confidently’ misunderstands the world, it won’t know its mistake until it receives new information.

Autopilot currently beeps to alert the driver when it senses a potentially dangerous situation unfolding, such as a vehicle cutting in – and could call on a remote operator if there is no driver in the car – but this certainly won’t catch all scenarios.

Sticking to the easy roads

However, Autopilot can still be judged by its past performance, just as regulators will do. For example, Tesla stores data of driver interventions, when the driver takes over; presumably because the car is taking the wrong action. By overlaying these on a map, we can see the roads where Autopilot performs well, and where it doesn’t.

Currently, Autopilot mainly performs well on highways, because although the speeds are higher, the amount of ‘chaos’ is much smaller. Importantly, there is the general understanding in many countries that if somebody stands in the middle of the highway, there is a good chance they will die. This is significantly different to city streets where pedestrians will often cross the road and expect other road users to avoid them.

What services can we build before Autopilot is ‘finished’?

So now we have a broad understanding of where Autopilot performs well, the risks of failure in different scenarios and how the driving behaviour should change according to whether there is a human in the car. Let’s look at how we can design a sequence of different services that become practical as the maturity and accuracy of Autopilot improves.

v1: One-way car hire

In the near future, Tesla car rentals could be offered from locations that are near to highways, e.g. “Park and Ride” stations, some supercharger sites or petrol stations. In some ways, this approach is quite similar to the approach taken by many human drivers who offer intercity rideshares:  drivers will often suggest a pickup point at the edge of the city in order to avoid getting stuck in traffic. So yes, you might actually get a human-driven taxi to go meet your ‘robotaxi’!

Of course, existing car hire companies offer this, but try getting a quote for a 3-hour one way rental between two small cities – they charge you a small fortune! Tesla could offer this service at a much more affordable price, because after you are done with the vehicle, it can autonomously drive off to where it is needed next.

They only need to charge enough to cover the vehicle cost of that additional driving, not a human to drive it – and in electric vehicles, that’s cheap. (In Tesla vehicles, that’s especially cheap, because they are engineering their vehicles to last a million miles while costing only fractionally more to manufacture.)

While the vehicle is empty, it’s fine for Autopilot to be a little more cautious, which will decrease the frequency of accidents and therefore reduce insurance costs. While you’re in the vehicle, you would be responsible for monitoring Autopilot and intervening as necessary. If there are any accidents, Tesla can use all of the sensor and camera data to see exactly what happened and take appropriate action if the customer wasn’t careful.

Charging the vehicles could be achieved in various ways, the low-tech option being simply to pay somebody to plug them into the Supercharger, and/or ask customers to do so. A hybrid approach could work well on sites where there are people permanently nearby, e.g. at motorway services. At those sites, customers could be asked to plug and unplug cables, but if no customers are around or arriving soon, an employee working nearby (such as a security guard) could be paid to do it.

A higher-tech solution is to have a robot arm to connect the charger. It’s not particularly difficult – Tesla displayed a prototype ‘snake charger’ back in 2015 – but it would probably be rolled out to guarded locations first in order to reduce the cost of vandalism, as they will inevitably be more fragile and expensive than regular Superchargers. They’re probably more attractive to vandals too, although perhaps Tesla will program them to fight back when attacked?

v2: Highway robotaxis

As above, but without needing the occupant to have a driving license. Autopilot would need to be safe enough that the occupants can be safely transported without any need for a human to intervene. That means children and others without a driver’s license could use the service.

However, occupants might find it reassuring if they’re able to slam on the brakes, even if they’re not allowed to fully take control. That would cover those times when a human can perceive an obstacle that the Autopilot hasn’t correctly identified.

The occupants will probably tolerate Autopilot driving a little cautiously, but might get frustrated if it really stops making sufficient progress. Perhaps this could be overcome by allowing occupants to summon a remote operator to help the car out.

v3: Robotaxi ranks

As Autopilot improves, pick-up points can gradually move further into the city. However, finding a safe place to stop is a tricky judgement call, and Tesla would be wise not to get a reputation for parking unsafely or blocking traffic.

Instead, robotaxis are more likely to first be available to stop in predetermined locations in a city, such as car parks, taxi ranks and designated pickup points near hotels and other buildings. This means that it’s still not always a door-to-door service, but as the number of stopping points increases, it becomes gradually more convenient.

v4: Full point-to-point robotaxi service

The final stage – the vision – is to offer the same service we’d expect from a human-driven taxi. Customers might be asked to identify a suitable and legal pick-up location when they book their ride, which might initially be approved by a human operator. Similarly for the drop-off location, a human operator might initially make a judgement call about the closest safe location to where the customer would ideally like to be dropped. Eventually this might be automated, but given the complex and chaotic nature of many city-centre environments, humans might still be involved for a while yet.

Taxi drivers won’t be the first out of a job

So taxi drivers can relax a little bit – your job is probably safe for another few years to come. It’s all the coach drivers, minibus drivers and perhaps even railway staff that should consider retraining for another profession. When I can rent a comfortable Tesla from one city and drop it off in another for the price of a train ticket, with no changing trains or waiting on cold platforms, long-distance public transport doesn’t seem so attractive any more.

And of course, car rental firms will be quickly pushed out of the market too. I foresee them trying desperately to squeeze revenue out of their old-fashioned gasmobiles by specialising in low-mileage rentals to customers who want to rent from a city location – and return it to the same spot. That’s a pretty small niche.

Their demise will be compounded by the rapidly diminishing resale value of fossil cars. Second-hand sales will fall off a cliff due both due to the increasing affordability and attractiveness of electric vehicles and because as more attractive mobility options become available, including those described above, fewer people will want to own their own car.

One day, car rental firms might be able to buy autonomous electric cars from other manufacturers, but given that Tesla is at least 5 years ahead of the big car manufacturers, we can expect to see major car rental firms going bankrupt before suitable vehicles hit the market.

Finding the golden business opportunities

Tesla is years ahead of the competition in developing efficient, reliable, comfortable autonomous electric vehicles. They have hundreds of thousands of suitable cars already on the road, ready to be turned into robotaxis when the software is ready. As Elon Musk puts it, “The fleet wakes up with an over the air update; that’s all it takes.”

Still it’s not clear what exactly Tesla has in mind for its “Tesla Network”. They have shown a prototype that looks very similar to the Uber app, but if they wait until Autopilot is good enough to replace taxi services in city centres, they will have missed years of first-mover advantage in other areas.

If they roll out the services described above, however, they will quickly become a force to be reckoned with in the intercity transport market. Doing so will demonstrate that their autonomous vehicles can be a profitable investment rather than just a luxury expense. It will help investors to maintain their belief in the long term robotaxi vision, even as the years inevitably pass by.

How can we get involved in the robotaxi revolution?

Tesla’s approach is highly vertically integrated: they do everything from making their own seats, to running their charging network and doing vehicle maintenance. So it’s unlikely that there will be much opportunity for us to get involved, other than working for Tesla, buying shares or investing in vehicles.

However, I hope that this article has given you a clear understanding of how AI technology can be used to create valuable business opportunities, years before the ambitious technical vision is achieved.

If you think about the industry you’re in, how could you use this same approach to discover disruptive opportunities that you can implement with today’s imperfect AI?