Simulation will drive Autonomous Driving
22 October, 2017
A small startup from Rehovot, Israel, believes that only massive parallel simulations can train Autonomous Vehicle AI systems in order to bring them into a Human Level. NVIDIA adopted the idea
A small startup company from Rehovot (near Tel-aviv), Israel, is suggesting a paradigm shift in the process of training the Artificial Intelligence systems of the soon to come Autonomous Vehicles. The idea of Cognata is simple: AI system is a virtual environment that receives digital inputs from the outside world. To get those inputs you don’t need to send the car to the streets, it is enough to generate virtual inputs, condition they are very accurate and represent the real life inputs.
In other words: training by simulating the real environment. Cognata is using computer vision and deep learning algorithms to automatically generate a whole city-simulator including buildings, roads , lane marks, traffic signs and even trees and bushes. It also adds a dynamic layer of realistic traffic model of other vehicles and pedestrians, representing their behavior. Even historic local weather conditions and lighting are added to stress test the system. Cognata also produce the relevant inputs that the vehicle receives from its sensors: cameras, radars, LiDARs and other sensors.
Cognata’s virtual reality simulator and engine enable autonomous car manufacturers to run thousands of different scenarios based on various geographic locations and driver behaviors. This vision was acknowledged by NVIDIA, who named Cognata Best Young Company in Israel at NVIDIA’s GPU Technology Conference (GTC Israel), held in Israel last week. Presented by NVIDIA CEO Jensen Huang, the award recognizes Cognata as the first among Israel’s hottest startups.
The CEO an founder of Cognata, Danny Atsmon, told Techtime, that above all there is also the Cloud layer that allows manufacturers to run heavy simulations. “According to RAND Corporation an Automomus Vehicle must drive at least 11 billion miles in order to achieve human capability, which defined as 1.9 fatal casualties for 100 million miles. For a standard machine learning training methods you need hundred of years to reach the goal. But with parallel simulation enabled by the cloud, you can do it quickly, in a practical manner.”
Cognata was established in 2016 and raised $5 million from investors. Today it expands its workforce and already sells its simulation services to car manufacturers around the globe.