Simulation needed for AI systems to drive a billion miles

With autonomous vehicle artificial intelligence (AI) systems needing around 11 billion learning miles before they can be deemed safe, companies developing selfdriving vehicles have realized that simulation holds the key to bringing autonomous vehicles to market any time soon. In addition to doing the necessary training, simulation allows developers to test edge cases — the unlikely and unexpected scenarios that pop up while driving — without the risk of a physical accident. The development of simulation systems has been given a boost by the partnering of Dassault Systèmes of France and Israeli company Cognata to embed Cognata’s Autonomous Vehicle Simulation Suite into Dassault Systèmes’ 3DEXPERIENCE platform. The partnership provides a first-of-itskind solution for autonomous vehicle makers to define, test and experience autonomous driving throughout the development cycle.

“Simulation is key at all stages of cyber systems engineering. Billions of miles must be virtually run before a car can be considered safe. AI-powered experiences that combine vehicle behavior, sensors and traffic models allow alternative designs to be tested in the concept phase to identify the optimal engineering solution,” said Philippe Laufer, CATIA CEO. “Integrating Cognata’s accurate and comprehensive offering into the 3DEXPERIENCE platform creates a unique solution to help our customers greatly reduce the time to market of safe autonomous vehicles, he added.”

“The partnership with Dassault Systèmes will hasten the development of autonomous vehicles, by making simulation an integral, seamless component of the engineering process,” says Danny Atsmon, CEO and Founder of Cognata. “The earlier simulation is utilized, the easier it is for engineers to modify each component of the autonomous vehicle and test it through a virtual environment, to see how it works once incorporated in the vehicle and confronted with unexpected edge cases.”

Dassault Systèmes’ 3DEXPERIENCE platform provides industry solution experiences such as smart, safe and connected to the transportation and mobility industry that transform the way next generation vehicles are designed, produced, delivered and operated. Cognata’s full product simulation solution leverages deep learning to enable autonomous vehicle manufacturers to run thousands of different scenarios based on various geographic locations, traffic patterns, and weather conditions.

By incorporating the Cognata simulation suite into the 3DEXPERIENCE platform and leveraging CATIA best in class systems engineering roles and applications, the two companies say they can now deliver a onestop- shop, outstanding environment to engineers for accelerated autonomous vehicle design, engineering, simulation and program management. Cognata recently raised US$18.5 million in a funding round led by Scale Venture Partners to finance the development of its technology.

Atsmon says the company will use the money to double its staff from the current 28, and expand commercial operations. Cognata wants to increase its international presence, specifically into the U.S., Germany, China and Japan, he added. Existing investors Emerge, Maniv Mobility and Airbus Ventures, as well as newcomer Global IoT Technology Ventures, participated in the round. Automotive majors like Audi are partnering with the company. In June 2018, Audi’s wholly owned subsidiary, Autonomous Intelligent Driving (AID), selected Cognata as its autonomous vehicle simulation partner in a multi-year agreement.

Cognata recreates cities from around the world, allowing an expanded range of testing scenarios, including AI-based traffic models simulating real-world traffic conditions. The simulation engine reproduces sensor input by emulating the specific sensors’ interactions with real-world materials. Autonomous Intelligent Driving has a fleet of test vehicles running the latest version of AID self-driving software.

AID develops the full software stack from AI and Machine Learning for perception and prediction to localization, trajectory planning and interface to sensors and computers. Initially focused on urban environment and mobility services, the AID software will eventually be a standard platform for all vehicles. Cognata is headquartered in Rehovot, close to the Weizmann Institute of Science. “We are extremely honored to have been selected by AID, a leading player in the autonomous vehicles space.

Thorough simulation is a critical and integral part of safely preparing autonomous vehicles for the road, and we look forward to a long-term collaboration with AID. The combination of the groundbreaking work done by AID and the end-to-end simulation offering from Cognata will safely accelerate commercial deployment of AID-enabled autonomous vehicles,” says the company.

Automotive Industries (AI) asked Atsmon why machines need to learn from machines. Atsmon: Every autonomous vehicle developer faces the same challenge — it is really hard to generate the numerous edge cases and the wide variety of real-world environments. Rand Corp estimates that, in order to get to a human level with high confidence, autonomous vehicles need to validate their software for 11 billion miles. This would take hundreds of years in real life. Our simulation platform rapidly pumps out large volumes of rich training data for the AI algorithms.

AI: How time will simulation save? Atsmon: Cognata’s simulated testing and evaluation environment shaves years off the validation time, and eliminates the safety concerns, high costs, and limited scalability of roadtesting in the physical world. Our scalable cloud layer runs millions of instances of its simulation software simultaneously to provide full coverage in only a few weeks. AI: What can you tell us about your new product? Atsmon: We are working on launching several products that we cannot expose yet with top tier OEMs, and are also working on integrating our software with partners such as Dassault systems and others. AI: How important will your technology be in introducing self-driving technologies faster and more cost-effectively? Atsmon: Self-driving is composed of a cutting edge set of sensors, maps, AI and processing technologies. Cognata is providing the proving ground to stress test, and to enable a fast development, verification and validation cycle. AI: Are OEMs other than Audi showing interest? Atsmon: Cognata is in touch with large number of OEMs worldwide. Due to confidentiality we cannot disclose the names of our clients.

Hi-fidelity Datasets for training AI “brains”

High-fidelity image datasets for training and certifying the artificial intelligence (AI) “brains” of autonomous vehicles earned Israeli start-up TheWhollySee the Inception Award at the 2018 NVIDIA GPU Technology Conference. The company, which received a US$100,000 grant and a Nvidia DGX Station personal AI supercomputer, has two full-time and two part-time employees.

Founder Dan Yanson said: “What makes this win special is that the judges recognized the value and potential in us. The prize money will give us a big push but the DGX workstation will lead to significant acceleration and dramatic impact on our activity.

“Despite being a small country, Israel is a huge force in the artificial intelligence industry,” said Bill Dally, the head of Nvidia’s 200-strong international research center when he made the award. TheWhollySee STRADA automatically augments realworld data to generate high-fidelity, high-diversity mixedreality data that does not require human labelling or postprocessing.

“Our image augmentation technology can generate annotated datasets of real-world imagery with automatic instance-level object segmentation in complex multi-object scenes. Edge cases and various object combinations can be specified as augmentation scenarios,” says Yanson. The company says the environmental perception by autonomous vehicles is powered by deep neural networks that need hundreds of thousands – if not millions – of examples to learn the appearance of common and not-so-common objects that they could encounter in real-world traffic. For supervised machine learning, one needs masses of annotated or labelled data, in which images are categorized by object labels and the precise pixels occupied by those objects. The safety of autonomous agents is predicated on their flawless environment sensing and perception, which require verification and certification using their sensor-specific data. However, the scene diversity and variety of edge cases achievable with physically driveable mileage are insufficient for fully autonomous driving certification.