Scaling Autonomous Vehicles with Machine Learning — Our Investment in Wayve

From a 2015 email to a seed investment: How Wayve uses novel machine learning to bring autonomous vehicles to the masses.

TL;DR — Compound led Wayve’s seed round. Wayve is an autonomous vehicle company utilizing a unique machine learning approach to drive cars autonomously, with little data, in novel environments.

In 2015 I read a research paper on semantic segmentation for scene understanding called SegNet. I sent one of the authors, Alex Kendall, an email congratulating him on his work and asking about commercialization plans.

Not my best cold email.

Alex said he was excited to get it running in a car soon, but wasn’t sure how.

Fast forward to 2017 when our Board Partner Drew Gray told me about a team of impressive University of Cambridge PhDs working on a new approach utilizing bleeding-edge machine learning techniques to get cars to drive themselves in novel environments.

Upon hearing one of their names I quickly searched my inbox and my email about SegNet from years prior popped up. I bumped the thread and re-met Alex and Amar.

Within an hour of meeting both of them, I realized that they were a rare breed of co-founders; both highly technical, but also incredibly entrepreneurial with a deep desire to commercialize and build a real, big business. After digging deeper into their research, it was clear that what they were building was special. Having seen 20+ AV startups in the prior 2 years, Wayve had by far the most differentiated technical approach I had ever seen. We led the seed round and I joined the board along with Gabe Matuschka at Fly Ventures.

Why do we need another autonomous vehicle startup?

In my time spent in the AV space I’ve repeatedly heard about the ceilings that various teams have hit along the way. While early teams at 510 Systems, which eventually became Waymo, saw rapid progress to ~80% autonomy, it was only after years of work they realized they needed to essentially start over from a technical perspective to reach 90%+ autonomy.

2017 Disengagement report — Disengagements per 1,000 miles.
Autonomous Vehicle Disengagement Trends & Statistics

We’re now starting to see this again with slowing progress in disengagement rates (an admittedly flawed metric) and continued failures across multiple major players. The top AV companies today drive in a few cities at most, with $100k+ of sensors, and plenty of existing maps and training data in those cities. And yet, few are on a realistic path towards scalable autonomy.

Today, the majority of companies utilize deep learning for computer vision (the perception layer) to allow the car to understand what it sees. The more annotated data, the better. They then use a rules-based approach to make path planning and driving driving actions that the vehicle takes once it understands what it sees. This results in the need to anticipate an almost infinite number of edge-cases across each layer of the self-driving stack. Where humans can generally understand and drive in completely new environments or situations, with this rules-based approach, today’s autonomous vehicles can not generalize nearly as well.

The traditional rules-based robotics approach is not the answer.

When looking back at the past two years of autonomous vehicle engineering I concluded that if we are going to truly bring AVs to the masses and break through the latest ceilings, we must use more advanced forms of machine learning.

Wayve is building truly scalable autonomy by innovating on the bleeding edge of bayesian deep learning (BDL), reinforcement learning (RL) , and more.

Deep Learning vs. Bayesian Deep Learning — via Concrete Problems for Autonomous Safety

By utilizing BDL and RL in an end-to-end model, Wayve is able to estimate uncertainty and propagate actions based on the entire AV framework, including the prediction and decision layers of the vehicle, greatly increasing safety. In AVs, the “state” refers to building out a representation of what is going on (and thus needed to drive), such as where cars and pedestrians are located, what the traffic lights are doing, and other known variables. Most teams hardcode this state of the vehicle today based off of rules (as previously mentioned), and if a given state of the vehicle doesn’t match the hand-coded rules, the vehicle can’t drive. This is why when many vehicles reach situations they haven’t seen before, they stop.

Wayve utilizes machine learning to learn the state of the vehicle and make a decision from that state to scale to unknown and complex scenarios. This enables a variety of decision making and data efficiencies, including the understanding of multi-agent decision making (i.e. complicated scenarios that arise where multiple vehicles are impacting each other, like a roundabout).

Put simplistically, Wayve utilizes a unique end-to-end machine learning approach to drive cars autonomously, with little data, in novel environments. This also means that their software enables a car to drive itself safely using only understanding of what it can see, just like humans do. And as of today, their team of incredible PhDs, AI/ML experts, roboticists, and software engineers have been able to do it with less than $5k in sensors per vehicle, on public UK roads.

Wayve’s sensor stack is largely made up of cameras

While the AV industry is still in its infancy, the early results from the Wayve team have been incredibly promising. I couldn’t be more excited to be along for the proverbial + literal ride, and for the team to pull back the curtain a bit on what they’ve been working on over the past year.

If you’re interested in joining one of the best self-driving teams in the world, Wayve is hiring!