Self-driving cars seem to always be on the cusp of being accepted by the institutions that govern our roads. We’re always on the edge of our seat waiting to hear if the latest test has been an outstanding success, or whether there’s been an accident that may bring the drive towards recognition to a standstill again.
A startup from founded by a team from Cambridge University Engineering Department is taking a different approach to making a self-driving vehicle, and they’re driving it with deep learning. Backed by a host of venture capital firms, such as Compound, Fly Ventures, and Firstminute Capital, the company is currently building end-to-end machine learning algorithms in order to make autonomous vehicles a reality.
Autonomous cars from big companies such as Google are at a point at which they are good, but not good enough for common use. The difference between Wayve’s methods and most self-driving cars being developed elsewhere is that other autonomous vehicles make use of a host of cameras and sensors, along with mapping tools and plenty of computer programming. But what is needed is a smarter brain in the vehicle, not more sensors, maps or programming.
Wayve goes beyond the other technique because they deem that such cars are not yet smart enough to handle the differing conditions present on an average road. They’ve even posted a YouTube demonstration video outlining how their technology works, showing a real car using their AI to self-drive on a real road.
In A Twizy
The Wayve video presented a Renault Twizy with a single camera and the gas, brake and steering control gear hooked up to a graphics processor and a computer running the reinforcement learning algorithms that they themselves developed, although this is a data job that can also be outsourced as well.
They had a human driver on-board to deliver input on what would be the best result. The computer was told that the optimal outcome would be for the car to move forward along a road without leaving the road, so the human driver/passenger would direct the car to go the right way, and then allow the computer to take over.
If the Renault came close to going off the road, the driver would stop it, get the car aligned and then ave the computer take over again. It took about 20 minutes for the algorithm to learn how to propel the car correctly, and after that it was able to continue on indefinitely.
Learning in the Deep
The team at Wayve believes the right approach is to use reinforcement learning algorithms. Rather than have a hand-engineered solution with rule-based systems, they aim to build data-driven machine learning for every layer of their system. This means the computer would learn from experience and not simply by given if-else statements.
The algorithm is similar to DeepMind AI, which has shown that deep reinforcement learning methods can lead to high-level performance in games such as Go, Chess and computer games. This process almost always outperforms any rule based system.
This actually creates a self-driving car that builds its own entire model of how to operate in the world, much the same way as a human driver has to, and this could be what it takes to bring autonomous vehicles into the mainstream.