How Cruise is testing its self-driving cars after the pandemic forced its vehicle fleet off California roads
SAN FRANCISCO — Autonomous vehicles, robots, and complex medical devices all require significant amounts of testing before they can be introduced to the general public, a timely and expensive process that can take years — if not decades — to complete.
To bypass that, companies are increasingly looking to simulation to test out complex machines in the virtual world before going to wide-scale production. Nvidia, for example, runs millions of simulations to perfect its robots that the chipmaker is hoping can help individuals with disabilities perform the most basic day-to-day functions — like cooking.
At Cruise, the self-driving arm of General Motors, those efforts are led by Tom Boyd — a former video game-maker who helped develop the popular titles "Sims 4" and "Star Wars: The Old Republic."
And the simulations are going to get even more important after Cruise and other autonomous vehicle (AV) firms halted on-road testing in cities like San Francisco and Pittsburgh due to the coronavirus pandemic.
"It's way easier to test the physical world in simulation if you can model it accurately," Boyd told Business Insider in an exclusive interview. "Before our car hits the pavement, it's already driven thousands of miles."
'Torture and torment'
The technology helps to solve one of the biggest hurdles in making self-driving cars ubiquitous on the road: training AVs to respond to dangerous threats. The vehicles need this training, but companies are hesitant to put the cars in real-world dangerous situations, for obvious reasons.
Simulation allows Cruise and others to "torture and torment" the machines in the virtual world to make sure the response is safe and appropriate when they go on physical roads, according to Boyd.
"We can use that to see what the reactions are, change the code. And we can even use simulation to create data that they can use to learn from so they can get better at changing their response to these situations," he said.
Getting the hundreds of millions of miles needed to ensure self-driving cars are safe on physical roads would be a monumental challenge in itself, besides taking decades to achieve.
But simulation allows companies to overcome that hurdle and bolster the on-road component with virtual testing.
Cruise, for example, says it runs 30,000 simulations per day, which is relatively akin to running 30,000 tests on the road.
And many of those tests are working on solving issues that may be months or years ahead of real-world scenarios.
"It's fundamental for most AV companies to start simulation the day they start the company," Boyd said.
The four pillars
At its most basic, simulation allowed Cruise to replay scenarios that its self-driving cars encountered on drives, right when they returned to the garage.
During on-road testing, the vehicles continually gather information about many aspects of the ride, including the surrounding areas and how quickly the AV is moving through certain neighborhoods.
Cruise can take that information, highlight areas where it thinks the vehicle may have underperformed, and run the exact same route in the virtual world to make improvements to the software.
But the company is often unable to test its vehicles in more dangerous situations. Simulation doesn't have that problem.
"We want to ask those 'what if' questions, because they are going to happen," said Boyd. "We can't guarantee that the world is behaving well, we can only control that our cars are behaving well."
For example, Cruise runs hundreds of simulations to test an unprotected left turn — when a stop light does not signal when a vehicle should go. Such testing would typically take thousands of hours in the physical world.
Those two encompass many of the simulations Cruise is currently running. But it can also create whole virtual cities that mirror the physical world — in Cruise's case, San Francisco, since that's where it does the bulk of testing.
The company can then introduce all kinds of threats, from erratic pedestrian behavior to out-of-control drivers, and see how their machines respond. But it's expensive.
"You're running this simulation, you're renting a lot of cloud compute, and you want to make sure that you've set that situation up right," said Boyd.
"That's the hardest one. You have to get your simulation looking very real and now you can use it to generate training data that would be hard to capture in the real world," Boyd said.
Aiming for comfort
Many of those tests are aimed at making sure the vehicle is safe. But Cruise is also focused on ensuring passengers are comfortable in the cars — another goal that can be measured by simulation.
One key metric that can help indicate whether a passenger will be at ease during the drive is acceleration.
"We don't know that we are going tens of thousands of miles an hour around the sun because we just don't feel that, we aren't changing that. But as soon as you change, you get the accelerations," said Boyd. "If you're in a car that was changing lanes poorly … you feel it and your body starts to move around."
Through simulation, Cruise can figure out scenarios where vehicles may be accelerating or decelerating at faster speeds — like hard braking — to learn the root cause and attempt to adjust the software to make those switches easier on passengers.
"There's this frequency range that human's inner ears and balance systems react poorly to," said Boyd. "We can measure the accelerations both directional, which is linear, and the angular acceleration, which is turning, and map them against what makes humans comfortable."
And while Cruise has safety and comfort as two of its overarching goals, figuring out what to test within those parameters can be difficult and requires close coordination between Boyd's team and the engineering team.
If the engineers are trying to test to make sure the car perceives oncoming traffic, "I don't want to be testing to make sure that my car's regenerative braking is exactly accurate," he said. "In this vast world of things you want to make real, you have to make the right choices."
Cruise's efforts on simulation show not just how powerful the virtual world can be in accelerating the pace of innovation, but also how technology can be implemented to overcome external factors that may inhibit other parts of the business.