The challenge for autonomous vehicles today is that they need lots of data to be able to navigate roads. This means that unknown roads can be a big challenge for the rides. Wayve says that its algorithm can drive on roads it has never seen using cameras and basic sat-nav.
The algorithm uses machine learning, something more and more autonomous cars are adopting. Wayve says that its system can drive like a human and needs no HD maps, no expensive sensor or compute suite hardware, no hand-coded rules, and can drive on roads it has never seen. The video below shows the car in testing in Cambridge, UK on public roads.
Wayve says that it doesn’t have to tell the car how to drive with hand-coded rules; everything is learned from data. The data-driven approach means that the vehicle can navigate complex, narrow urban European streets. The system learns end-to-end according to Wayve and learns like a human via computer vision. The system has imitation learning so it can copy behaviors of expert human drivers.
Reinforcement learning allows the vehicle to learn from each safety driver intervention. The model is capable of learning steering and acceleration input. The ability of the system to learn input features from input data most relevant to control makes for reduced sensor and compute cost. The team says its sensor and compute costs are 10% of that of traditional approaches. The video shows the car navigating with a consumer-grade GPS. Wayve notes that the system can navigate in traffic and in the rain.
Wayve has given no indication of when its autonomous system might be commercialized. The company does note that its test vehicles have a safety driver behind the wheel at all times.