How Google Cars learn to drive through simulation
AUTONOMOUS. Waymo, Google’s subsidiary in charge of the development of the autonomous car, has just published on its blog the strategy used to face new situations. In this case, it is a five-lane crossing with an arrow flashing to the left for the center lane, allowing the car to turn while other cars arrive on opposite lanes. This type of crossing is rare in Mountain View (California), on Google land, but quite common in Phoenix and its surroundings (Arizona) where Google Cars line up thousands of kilometers currently. And crossing is tricky for both a human and an autonomous vehicle. Indeed, it is necessary on the one hand to wait to cross the opposite lane at the right time without risking the accident while not dragging too much so as not to frustrate other motorists waiting behind to pass. “As with humans, the key to learning is training,” the Waymo team explains on the blog.
Before borrowing independently, a complex crossroads like this one, the Waymo car software trains on a simulator that reproduces the real world as best as possible.
VIRTUAL. The engineers then use a powerful simulator that is in fact a realistic virtual world in which they can reproduce the slightest kilometer traveled in the real world. Every day in this simulator, 25,000 Waymo self – driving cars travel about 13 million km-or 33 times the distance between the Earth and the Moon. Virtual cars thus test a large number of scenarios and practice maneuvers before this new knowledge is applied in the real world.
This training takes place in several stages. This is how it happens in the case of crossing with the flashing arrow.
1 – A very detailed view of the real world : Using numerous onboard sensors, Waymo engineers reproduce in the virtual world the most realistic possible version of the crossroads, respecting distances, lanes, curves and traffic lights.
2-Driving, driving and driving : Once the carrefour and its flashing light are well digitized, the software can repeat the scenario thousands of times and thus learn to react better. It also becomes faster and smarter because it is able to translate this learning to other similar hubs.
3-Create variations: From the simulation of this crossroads, Waymo diversifies the scenarios for example by changing the speed of the cars or the synchronization of the lights. The virtual autonomous car must adapt. Engineers can go further by increasing the density of traffic or by adding pedestrians, cyclists or motorcycles that cut the road or joggers that zigzag along the roadway. The idea is to see how this will affect the driving of the car.
4-Verify, validate and repeat: The new skill acquired by the software at this crossing is then integrated into the knowledge base of all self-driving cars in the Waymo fleet. The results of the simulation will therefore be repeated and verified in the real world.
According to Waymo, most of the improvements made on their driving software come from simulation. In this virtual environment, cars have traveled more than 4 billion kilometers.