Thank you for Subscribing to Auto Business Outlook Weekly Brief
Self-driving cars have the ability to dramatically improve road safety while also giving people the freedom of mobility. However, to ensure safe deployment in the near term, self-driving vehicles will require a unique approach to augment traditional methods of vehicle testing.
It’s impossible for current autonomous vehicle test fleets to encounter the rare and dangerous scenarios or extreme conditions necessary to train a self-driving car to safely handle the real world. For example, a small child wearing dark clothing running out in the street at night time. It’s not a common scenario that can be repeatedly tested, but it’s one that a self-driving car needs to be able to safely and properly react to. What’s more, current physical vehicle testing can’t validate at a large enough scale—it’s limited and a costly endeavor for companies working on AVs. “Through advances in artificial intelligence and accelerated computing, simulation has emerged as a way to safely test and validate self-driving technology before it takes to the streets” Through advances in artificial intelligence and accelerated computing, simulation has emerged as a way to safely test and validate self-driving technology before it takes to the streets. Leveraging a virtual AV test fleet in the cloud, operating on an open, bit-accurate solution can help enable the development and validation of AVs at scale without putting others on the road in harm’s way. This allows for greater efficiency, cost-effectiveness and most importantly, safety, than what’s achievable with real world test drives. There are a number of companies working on simulation for the automotive industry. At NVIDIA, we’ve developed a high-fidelity simulation tool that leverages the computing horsepower of two different servers, making it capable of generating billions of qualified miles of AV testing. The first server runs software to simulate a self-driving vehicle’s sensors (cameras, radar and Lidar). Powerful graphics processing units (GPUs) generate photoreal data streams that create myriad scenarios and environments—allowing testing for rare and difficult conditions like rain and snow, to different times, to changing road conditions. The second server contains a powerful AI-enabled computer that runs a full software stack that operates in the vehicle. It processes the simulated data as if it was coming from the sensors of a car actually on the road and sends driving commands back to the first server. By making this platform open, third party developers can provide countless world models, vehicle dynamics models and traffic scenarios. Being able to simulate the environment, traffic and vehicle behavior using dedicated models for each domain, AV tests can play out unscripted. This makes it possible to find the ‘unknown’ unknowns — the edge cases a self-driving car could encounter that may not be commonly encountered by human drivers. By mining these situations in simulation, self-driving vehicles can be rigorously tested over and over again, 24/7, to improve the hardware and software makeup of the car before its deployed. This open, scalable solution provides the way to accelerate the safe deployment of AVs.