Simulation Technology Driving the Autonomous Vehicle Training Market


Over the next several decades, autonomous vehicles (AVs) are likely to be a common sight on roadways both in the United States and around the world. But ensuring that these vehicles can safely operate without driver intervention on the various types of highways, streets, and roads, while simultaneously managing traffic, other vehicles, and changing weather conditions, requires significant testing to feed data into the driving algorithms.

A 2016 Rand Corporation study estimated that AVs would need to be test-driven 11 billion miles to demonstrate with 95% confidence and 80% power that their failure rate is 20% better

than the current human driver failure rate of 1.09 fatalities per 100 million miles. Given the current fleet of AVs in testing, Rand indicated it would take 500 years to complete this level of testing in the real world, on real roads. Currently, only about 11 million miles of AV driving have been completed, according to Heikki Laine, Senior Director of Strategy and Policy for Cognata, a 2-year old, Israel-based company focused on providing AV simulation technology to original equipment manufacturers (OEMs).

“I think the latest numbers, when we look at the big players of Waymo and Uber and Cruise, put them all combined at something like 11 million miles [of testing],” Laine says, noting that this total falls far short of the billions of miles that are generally accepted as a minimum amount of driving and training that will be required for roadworthy AVs. “Even with the bigger testing fleets, this is not an attainable number to do it physically on the road, and creates the demand for at-scale simulations for AV makers.”

Creating and Training in Virtual Environments

The market is not limited to one or two companies; Braiq, Naoh, Phantom AI, and Mighty AI are just a few of the firms involved in the space. They are poised to take advantage of OEMs’ desire to log additional training miles in a virtual environment over the next several years.

These companies use a combination of technologies, including deep learning, mapping, and sensor technology, to create virtual cities that mirror their real-life counterparts. In addition to mapping roads and other objects that are present in the real world, such as static objects (traffic control devices, buildings, and roadside items like hydrants and road barriers), these virtual environments can model weather conditions and changing circumstances that mimic real-life variability and the introduction of moving objects (people, bicycles, animals) and other vehicles. Localized data, such as driver behavioral data and traffic patterns, can also be incorporated into the simulation. The goal is to create a digital environment that accurately mimics the real world.

Automotive Simulation Services Expected to Take Off

Not surprisingly, the market for simulation services is projected to take off rapidly over the next several years, according to Tractica’s updated report, Artificial Intelligence for Automotive Applications, which forecasts that automotive simulation, categorized as “building generative models of the real world,” will see 2025 software revenue hit $929.7 million, with cumulative spending between 2017 and 2025 totaling $3.2 billion. More details can be found in the report, including regional breakouts, as well as segmentation for hardware, software, and services revenue.

The goal of automotive simulation is to use technology to train autonomous vehicle software before real vehicles ever hit the street, and augment the on-road testing, particularly with respect to training AVs to navigate and manage situations that would be too difficult or dangerous to test in the real world, such as having a pedestrian entering the path of travel of a moving vehicle.

“We don’t view simulation as a replacement for on-road testing or proving-ground testing,” Cognata’s Laine says. “I think the cases that are coming out and becoming more important for simulation are really the rare events and the events that are just too dangerous even in a proving ground to use with the human subjects. So, the big ones, if you look at where the governments are focusing on, most of these are relevant to pedestrians and bicycles.”

Furthermore, by testing AVs using simulations, it is possible to train vehicles to adjust to specific regions, and the specific road signs, traffic control devices, and even the typical driving styles of people in the region. For example, Cognata’s platform can quickly acquire local maps and traffic rules from some local jurisdictions and then use these elements to customize the simulation environment. However, variable elements need to be constructed over time, given the wide range of variability.

“Region-specific things, like driving behavior and so on, is something that we are building with AI,” Laine says. “This is proprietary to us, and this is created new for each region or geography and so on. But, there is a variety of places where we can get driver behavior and we can collect it, down to relatively accurate representations of what drivers do, and we are using AI to procedurally train our behavior models from that.”

Mimicking Scale and Accuracy in the Cloud

Perhaps the biggest challenge for OEMs using simulation is ensuring that the simulated worlds are as big and accurate as the real world. That is where Cognata says it has the ability to scale its platform in the cloud, allowing OEMs to incorporate real-world levels of vehicles, pedestrians, traffic control devices, and weather conditions, across a number of simulated environments. This allows significant training miles to be accumulated in near real-world conditions, without the cost and safety concerns of conducting tests in the real world. Given the steep challenge of racking up enough training mileage to ensure the safe operation of AVs, it is likely that autonomous simulation vendors will thrive over the next decade.

Comments are closed.