London-based, Wayve, A driverless car startup has built a machine learning model that can drive two different types of vehicles: A car and a van. This is the first time that the same AI driver has learned to drive multiple vehicles.
The news comes less than a year after Wayve showed it could take AI trained on the streets of London and use it to drive cars in four other UK cities that would normally require a major redesign.
“It’s like going somewhere new and renting a car. You can still drive”
Wayve CTO Jeff Hawke.
The bias suggests that Wayve’s approach to autonomous vehicles (AV), in which a deep-learning model is taught to drive from scratch, could help it grow faster than leading companies like Cruise, Waymo, and Tesla.
Wayve is a much smaller company than its larger and better-funded competitors. But it’s part of a new generation of startups, including Waabi and Ghost, also known as AV2.0, that are abandoning the robotic thinking adopted by the first wave of driverless car companies—where driverless cars rely on highly detailed 3D maps and separate modules for recognition and planning. Instead, these startups rely entirely on artificial intelligence to control their vehicles.
The robotics approach has brought robot taxis to some streets in Phoenix and San Francisco — but at great cost and with little sign that the services will soon spread beyond the pilot.
Wayve and others hope to change that by replicating what deep learning has done for computer vision and natural language processing in self-driving vehicles, allowing them to better adapt to unfamiliar streets and scenarios without the need to keep them updated or manual complex maps and software systems.
Wayve’s headquarters in London has a new Maxus e9 van, which is parked next to an existing Jaguar I-PACE fleet. The van has the same six webcam sensors as the cars, but they are placed higher and at a different angle.
This means that the model’s input—the video feed from each camera, which it looks at about 30 times a second—is different from vehicle to vehicle, but the AI has learned to control them from both perspectives. The AI also had to adapt to the larger size and mass of the van. It has a different turning circle and takes longer to stop.
A car and a van may have the same decision-maker behind the wheel, but those decisions must be made in different ways. Under the hood of the van is a mess of wires and usual computer parts that translate the commands of the model to the vehicle being driven.
Autonomous AV2.0 vehicles adapt to the concept of unknown road conditions . The new big idea is to make self-driving cars that can go anywhere
The conventional approach to driverless cars is slow and complicated. These startups believe that AI all-in will arrive faster.
Wayve’s AI Model
Wayve’s AI models are trained using reinforcement learning, where it learns through trial and error, and imitation learning, where it mimics the actions of human drivers.
Thousands of hours of driving data were required to train the model to drive. First, the company trained its AI model to drive the van in a simulation that took another 80 hours of data.
This surprised the team. “When we started this project, we didn’t know how much data would be needed to generalize the system,” says Becky Goldman, a Wayve researcher. But the result suggests that the model adapts to new vehicles faster than expected. Way also found that learning to drive the van improved its performance in the car.”
When the model could drive both a van and a car in the simulation, Wayve took it on the road. Wayve security operator Naomi Standard sits in the vehicles as they drive themselves.
He admits he was scared the first time he drove the van: “I once had a driving instructor when I first took the driver out.” But the van coped well with London’s narrow streets, roadworks, pedestrian crossings, buses and double-parked cars.
Wayve’s performance at Ghost in Mountain View, California, was impressed with Wayve’s performance and agrees with the company’s overall sentiment. “Robotics is not the way to do it,” says Gierak. But he wasn’t sold on Wayve’s total commitment to deep learning. Instead of one large model, Ghost trains several hundred smaller models, each with their own specialty. It then programs simple rules that tell the self-driving system which models to use in which situations.
Ghost’s approach is similar to that of another AV2.0 company, Autobrains, in Israel. But Autobrains uses another layer of neural networks to learn the rules. According to Volkmar Uhlig, co-founder and CTO of Ghost, dividing the AI into several smaller parts, each with its own functions, makes detection easier. An autonomous vehicle is safe “Something happens at some point,” he says. And the judge asks you to show a code that says: “If there is a person in front of you, you must brake.” That code must be there. Code can still be learned, but it would be difficult to find in a large model like Wayve, says Uhlig.
However, the two companies pursue complementary goals: Ghost wants to make utility vehicles that can drive themselves on highways; Wayve wants to be the first company to put driverless cars in 100 cities.
Wayve is now working with UK grocery giants Asda and Ocado to collect data from their city delivery trucks.
However, many companies are far from market leaders.
Cruise and Waymo have amassed hundreds of hours of driverless time in their cars and are already offering robot taxis to the public in some areas.
