UK-based Wayve is the outsider in the autonomous driving industry with a unique solution that breaks all of the usual rules and could potentially upend the entire sector – but it’s an approach I continue to think is a long shot at producing a viable solution.
Wayve – which came out of Cambridge University’s AI research – has announced a partnership with online grocer Ocado and deliveries in London are due to commence this year albeit with a safety driver ready to take over in case of problems.
The company has also just raised a massive $200 million in a new round from Microsoft, Virgin and Baillie Gifford as the largest investors. This brings total funds raised to $244 million, highlighting just how large the step up in terms of scale that the company is about to take.
Wayve’s solution is unique in that its fleet of Jaguar iPace vehicles use nothing but cameras and a single deep learning model to run the entire system, from perceiving the environment to taking the driving decisions.
Almost everybody else uses a combination of Lidar, radar and cameras to perceive the environment and software to make the driving decisions, which I continue to think will ultimately prove to be the right choice.
This is because there are two big problems with Wayve’s approach:
1. Causal understanding
Deep neural networks like the one that Wayve is using appear to be far more able than they really are. This is because they are nothing more than very sophisticated pattern recognition systems that match the statistical characteristics of data to outcomes. This means that the system that pilots Wayve’s vehicle has no idea it is driving a vehicle, no understanding of the road and no knowledge of the rules that govern it.
In a purely static and finite environment such as a parking lot or a racetrack, this is not a problem because history is a 100% guide to the future. But the open road is neither of these two things. This means that when the vehicle meets a situation, lighting condition or event that it has not seen before, it will have no ability to adapt, and therefore will be unable to execute the correct manoeuvre. Furthermore, even the tiniest perturbations in the data set could cause the system to fail for no apparent reason.
However, Wayve says it has made progress on this front and claims to have demonstrated that it is able to make use of training data from London to teach its vehicles to drive in other cities. This is a sign of adaption and generalization that, if verified, indicates that Wayve is doing something unusual with its algorithms.
Deep learning systems are black boxes where the operator can see the output and the input but has no real idea how the machine drew the conclusion that it did. Any system where lives are at stake requires regulatory clearance, and unless Wayve can explain precisely how its system does what it does, then it will be very difficult to convince anyone that the system is safe.
This is a problem that has been vexing the entire AI industry, but by using software for the piloting decisions, the issue can be partially alleviated. I am not convinced that Wayve’s system was designed to deal with this issue, as I suspect that when the company was founded, the founders did not have a clear idea of whether the system would work at all.
Hence, a tricky problem of trying to work out what is in the black box and how it is working lies ahead before the company has a chance of obtaining regulatory approval.
This is a novel approach, but it is clear from the investors who have put in the most money that this is a high-risk venture. Virgin and Baillie Gifford are known for investing in long shots while Microsoft’s investment is arguably much lower risk. This is because Wayve, like all deep learning-based systems, needs a lot of data and compute, the latter of which is all going to come from Microsoft in form of Azure.
Hence, I suspect that, like Open AI, almost all of the money that Microsoft puts in will come back in the form of Azure revenues regardless of whether the technology succeeds or fails.
I think that Wayve remains a long shot. But, if I am to be proven wrong, and if it can solve the two problems mentioned above, it will be the entire industry that is turned upon its head.