Wayve uses machine learning algorithms to put self-driving cars on the road

Wayve, founded by a team from the Department of Engineering at the University of Cambridge, uses machine learning algorithms to enable self-driving cars to drive on unfamiliar roads using only cameras and basic satellite navigation.

Since NVIDIA released its end-to-end deep learning technology for self-driving cars in 2016, countless companies, organizations, and even enthusiasts have used this technology to make self-driving demos. The simple network structure can achieve direct mapping of camera input to brake, accelerator, and steering wheel output. However, this low threshold also means that it can only solve a few problems and is difficult to cope with the complexity of specific driving environments. Some experts even believe that end-to-end is not suitable for developing practical unmanned driving systems. It can be used for demos, but large-scale commercial use may be very difficult.

Is end-to-end only suitable for demos? Wayve, a self-driving software company founded by a team from Cambridge University, doesn’t think so. They don’t use high-precision maps, expensive sensors such as lidar, and of course, they don’t manually input rules into the car. With only 20 hours of training data, they can drive on roads they have never driven on.

The Wayve R&D team believes that since it is autonomous driving, there is no need to manually encode some regulations, and to fully demonstrate its intelligent characteristics. The team adopted the currently popular deep learning reinforcement learning algorithm to build an autonomous driving system that can slowly learn to drive like humans.

After iterating through the three steps of exploration, optimization, and evaluation, deep deterministic policy gradients (DDPG) are used to solve the lane keeping problem.

While existing image classification architectures have millions of parameters, the Wayve team’s network architecture is a deep network with 4 convolutional layers and 3 fully connected layers, with a total of less than 10,000 parameters, all of which are performed on the car’s GPU.

In the reinforcement learning simulation test, curved lanes, road textures and lane markings are randomly generated, and then the strategy is optimized based on the collected data and repeated over and over again.

An end-to-end zero-shot framework combining image translation and behavior cloning

While most self-driving car companies use simulation to validate their systems, Wayve lets self-driving cars learn extensively in simulation how to handle rare edge cases. Wayve trains its cars to drive in simulation and transfers what it learns to the real world.

Rather than treating simulation and the real world as two distinct domains, Wayve designed a framework that combines the two, making it possible to both train steering decisions in simulation and exhibit similar behavior in the real world without real demonstrations.

Wayve’s model consists of a pair of convolutional variational autoencoder-style networks originally used for image translation, namely Unsupervised Image-to-Image Translation Networks (UNIT). The model is able to translate between the two domains without any known alignment or correspondence between them. The figure below is an example that captures the main layout of the scene. It is worth noting that the visual fidelity of the simulator is not the most important when learning to drive. Their simulated world is like a cartoon, which can still do a good job of simulation. Wayve research claims that content fidelity is more important than visual fidelity. However, effectively simulating the behavior of other traffic participants remains a huge challenge.

Simulate scenarios based on real-world driving data and carefully designed edge cases

The car is driven by a model-based deep reinforcement learning system, which learns a predictive model from real data collected offline. This allows the model to learn and use data from new scenarios imagined by the predictive model to train driving.

Wayve is committed to developing richer and more powerful temporal prediction models and believes that this is the key to building intelligent and safe self-driving cars.

Currently, the system has been deployed on the Jaguar I-PACE car. This car won the title of European Car of the Year in 2019, and will collect data throughout the UK and continental Europe in the future. At present, as the data gradually accumulates, its driving algorithm may reach 95% of the quality of human drivers, and be able to handle traffic lights, roundabouts, intersections, etc.

Although some people may feel that end-to-end autonomous driving systems are neither smart nor flexible, and problems that occur are difficult to explain, Wayve is using its powerful algorithms to prove that this deep learning technology can not only be used for demos, but can also ensure safety and be used commercially in the future.