From dorm room to Series A: How and why Balderton invested in Wayve
tl;dr: I’m excited to be announcing our investment in the Series A of Wayve, alongside Eclipse Ventures. Wayve are a Cambridge-born, London-based pioneer focused on the intersection of robotics, machine learning, computer vision and reinforcement learning. Their first applications will be in self-driving cars and you can read the official take on our website.
Hang on, but who are Wayve?
I first met Wayve in a coffee shop in Cambridge back in 2016. At the time there were four founders, no specific product and they were not yet called Wayve. My partner James had actually already met some of team in their dorm room a year earlier. The team were coming to the end of PhDs and post-docs at Cambridge and were frustrated by what they saw as group-think in the way machine learning was being used across a variety of applications. Fast-forward three years and the company ultimately founded by two of the original team — Amar Shah and Alex Kendall — has focused on the specific application of self-driving cars and is now a team of over 40 with, yes, a name and logo.
But isn’t self-driving already solved?
In a word, no. When I look back at my notes from 2016, the phrase that sticks out is ’thinks it’s all done wrong today’. The founding team at Wayve have been unfailingly consistent in their view that many attempts at applying AI in industry were doomed to limited outcomes, if not outright failure. From this starting point, they hunted for a market where there was significant value and a shoal of fish swimming in one direction where they could be contrarian and found it in self-driving cars.
Despite billions of dollars of investment and ambitious declarations, the self-driving car industry is stuck. Valuations are being cut, investment is dropping and multiple players push back trials and launches. Wayve believes the problem is the technical approach used by those legacy players.
Stemming from successful contenders in DARPA’s Grand Challenges of the early 2000s, the prevailing strategy is to break the task of driving down into many component parts. These component parts are solved by sub-systems that often use forms of AI. Sensing immediate environment around the car becomes a LiDAR problem, lane-following a Computer Vision challenge, tactical route planning relies on bespoke, high-grade mapping and so on. Each of these sophisticated subsystems views the world through their own lens and sends back data to a central decision engine. This decision engine consumes the inputs and uses a long list of manually coded rules to figure out what to do next.
There are two huge problems with this approach. The first is universality, the second cost.
The Universality Challenge
A manually-configured, centralized decision engine struggles to deal with open-ended driving. The average human learns to drive in fifty hours and can then do a decent job of driving in the US, the UK, Japan and India. They are probably best in their own car, but with a little extra coaching, can also drive a bus or a truck. In contrast, state of the art self-driving requires millions of hours of training and yet yields algorithms that are limited to short, highly-prescribed routes that have most of their variables (traffic, weather, pedestrians, etc) controlled for them. All of those massive estimates of market size for self-driving cars? They don’t apply if you can only shuttle people a few miles on a fixed route in Detroit.
The Wayve team have bucked this trend since the day they started. Instead of multiple systems and a manually-coded core engine, they built an architecture on end-to-end deep-learning. And, much like a human, they rely on just vision and a simple 2D GPS map. The early results have been impressive — in just a day, reinforcement techniques allowed their car to follow lanes; with just 50 hours of training, their car could navigate dense, urban streets, a far cry from the Arizona freeways preferred by many legacy players.
The Cost Challenge
The reality is that the economics of self-driving are tough. There are questions around utilisation rates of any taxi (whether robo- or not) but another key concern is that the cost of the vehicles is massive. Brave predictions abound about where Moore’s Law will get that price point in five to ten years but, for now, you can easily add between $100k-150k per car to fully equip it with the appropriate sensors.
Again, Wayve’s approach changes things. By simplifying the sensor set required, the capital cost of the system is dramatically lower. In addition, the use of Reinforced Learning has created a more effective training loop — there is no need for costly rules engineers to sit in between the real world and the car, spotting and configuring around all conditions encountered. And that was before Wayve’s break-through on ’Sim2Real’ means they are able to train their model in simulation while most competitors use simulations only for testing.