No matter how hard you test a self-driving car, it's not enough. There are too many situations in which the system of artificial intelligence has to be trained, and too little test time on the highways. That's why Nvidia Drive has developed Constellation, a simulation platform that tests autonomous vehicles in virtual reality. Nvidia said the system will allow it to test the AI for self-driving cars by driving billions of miles in VR simulations.
Announced at the GTC event in San Jose, California, the technology is a fusion of two very different technologies Nvidia's GPUs: self-driving cars and VR. (We predicted this type of system at our GamesBeat Summit in 2017, when car software manufacturers used games to simulate self-driving car conditions.)
The cloud-based system for testing autonomous vehicles with photorealistic simulation can be a safer, more scalable method to get self-driving cars on the road. Nvidia CEO Jensen Huang said at the opening speech of GTC 201
"We at Nvidia are dedicated to solving this problem," Huang said onstage. It's just so important. It's a security problem. "
He reported last week that an accident occurred when a pedestrian died after being hit by a self-driving car in Arizona.
The first server operates Nvidia Drive Sim software to simulate the sensors of a self-driving vehicle B. Cameras, lidar and radar. The second contains a powerful Nvidia Drive Pegasus AI computer, which manages the entire autonomous vehicle software stack and processes the simulated data as if it came from the sensors of a car on the road.
"The introduction of self-driving cars requires a solution to test and validate billions of kilometers of driving to achieve the safety and reliability customers need "said Rob Csongor, vice president of automotive at Nvidia. "With Drive Constellation, we've achieved this by combining our expertise in the visual computing and data center industries with virtual simulation to increase the robustness of our algorithms by testing billions of miles of custom scenarios and rare case studies, and all in a fraction of the time and cost required on physical roads. "
The simulation server is powered by Nvidia GPUs, each of which generates a stream of simulated sensor data that is fed into the Drive Pegasus for processing.
Driving commands from Drive Pegasus are fed back to the simulator and complete the digital feedback loop. This "hardware-in-the-loop" cycle, which occurs 30 times per second, is used to validate that Pegasus's running algorithms and software are operating the simulated vehicle correctly.
The Drive Sim software generates photorealistic data streams for a wide variety of test environments. It can simulate different weather conditions such as rainstorms and snowstorms; Glare at different times of day and limited visibility at night; and all different types of road surfaces and terrain. Dangerous situations can be simulated in the simulation to test the responsiveness of the autonomous car without ever putting anyone at risk.
"Autonomous vehicles need to be developed with a system that ranges from training to testing to driving," said Luca De Ambroggi, Research and Analyst Director at IHS Markitk, in a statement. "Nvidia's end-to-end platform is the right approach – Drive Constellation for virtual testing and validation brings us one step closer to producing self-driving cars."
Drive Constellation will be available to partners with early access in the third quarter of 2018 be available.
Huang said the neural network needs to be properly trained, and that requires a continuous platform, from data acquisition through sensors to driving.
He said the US has 770 accidents per billion miles. But a fleet of 20 test cars can only cover one million miles a year. That's not enough to cover all sorts of problems that can cause accidents.