Playing God (virtually) with self-driving vehicles
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Rebecca Zisser/Axios
Self-driving test vehicles can’t possibly log enough real-world miles to ensure they are safe under every circumstance, but new tools allow manufacturers to test more of the rare but still possible scenarios they might encounter.
Why it matters: Extensive simulation is critical if autonomous vehicles are to be ready any time soon. Real-world testing on public roads is slow and limited — AVs could drive around for decades waiting for challenging conditions or interesting anomalies that would help train them to handle those so-called edge cases.
What’s happening: Nvidia this week released its new Drive Constellation simulation platform, which allows developers to play with variables like the angle of the sun, types of street lights, weather, road conditions, traffic and pedestrian behaviors.
- Rather than having a fleet of 50 or 100 AVs on the road, developers can test thousands on a virtual proving ground at the same time.
- Toyota Research Institute-Advanced Development is the first customer. It plans to simulate the equivalent of billions of miles of driving in challenging scenarios.
How it works: The cloud-based platform contains two servers operating side-by-side — one simulates sensor output from the car and the other receives that data, makes decisions, and then sends commands back to the simulated vehicle.
- Developers can create all kinds of scenarios — for example, an AV reacting to another car cutting into its lane in heavy traffic during a foggy night on wet roads.
- They can run the same test under other conditions simultaneously, completing months or years of testing in a fraction of the time.
- During a demonstration at this week's Nvidia GTC conference, I joked with a company engineer that he was playing God each time he hit a switch to change the driving conditions.
Waymo has driven more miles than anyone — 10 million on real roads and 7 billion on virtual roads, using its own simulation technology — but even that's not enough, based on research by the Rand Corporation, a policy think tank.
- AVs would have to be driven hundreds of millions of miles and sometimes hundreds of billions of miles to demonstrate their reliability in terms of fatalities and injuries, per Rand.
- Under even aggressive testing assumptions, it would take tens and sometimes hundreds of years to drive these miles with existing fleets.
- Developers "cannot simply drive their way to safety," the report found.
I heard a lot about simulation this week in Silicon Valley from all the major AV companies.
- "You have to lean into simulation to create these scenarios you never saw in the real world — but you might have," Sterling Anderson, co-founder of Aurora Innovation, told me.
- Simulation also gives developers the confidence to tackle edge cases that are too dangerous to test in the real world.
The catch: The simulation has to be a faithful representation of what would have happened in the actual vehicle. If the timing is off by even a micro-second, it can change the outcome, Anderson said.
The bottom line: Real-world testing still matters, as I learned riding around San Francisco and Silicon Valley this week in a variety of test AVs. Every on-road encounter unlocks new scenarios for developers to manipulate virtually in search of the perfect driver.
