A Conversation with Kyle Vogt, CEO of Cruise Automation
What gives General Motors an advantage over other companies that are developing autonomous vehicles?
Removing the driver from the vehicle is a monumental engineering challenge. It’s not one you can solve by bolting technology onto an existing vehicle. Rather, it requires reaching into the core of that vehicle and redesigning electrical and system architecture to ensure a level of redundancy that keeps a vehicle safe no matter what.
That’s a hard thing for many companies in this space to do. But GM understands how to conduct crash testing and build safe vehicles—because they’ve been doing it for 100 years. Bringing that expertise to the table and combining it with the state-of-the-art AV technology creates a unique skillset that’s really hard to replicate.
How does Cruise choose where to test its autonomous test vehicles?
I think about this in terms of extremes. If we put ourselves in the most complex environments first, we can make sure that our software is not overfitted to any particular environment. GM is currently testing in multiple markets, and we are the only company in the world that tests regularly in the complex urban environment of San Francisco.
This has given us valuable perspective on the relative complexity and challenge of driving in dense urban environments versus more straightforward and simple suburban environments. For example, if we look at the numbers, we see that the most challenging maneuvers for an autonomous vehicle, like dealing with construction or having to drive into an opposing lane to pass another vehicle, happen 20 to 50 times more often in complex environments. That’s why basing the maturity of a project on an incident rate recorded only in a simple environment can be misleading. We’re trying to be transparent and realistic about the real-world situations AVs may face.
Chevrolet Bolt with Cruise Automation.
Are you using methods other than road testing to improve vehicle performance?
Road testing is obviously important, but deep simulation capability is also playing a key role. We have a number of different simulators that exercise various pieces of the autonomous vehicle software stack. At any given time, there are 150 or more virtual self-driving cars in simulation. These cars are running through scenarios that we have observed during road tests that have either given a vehicle trouble or that we have identified as a potential safety issue needing verification. When coupled with on-road testing, simulation is a powerful tool to achieve, and then demonstrate, the level of performance required for launch of a commercial product.
What capabilities has GM added to its toolbox to support production of AVs?
Mapping is a core technology that we have developed in-house. We saw a gap in the market for mapping, and in response, built a technology that has turned into a valuable asset. Sensing is another one of the current bottlenecks in AV performance. With our acquisition of Strobe, we’ve taken some of the best sensing technology on the market and accelerated its roadmap to scaled deployment. We see a path to take out 99 percent of the cost of this new technology, which removes a critical barrier to our progress.
What role does deep learning play in the engineering process?
Deep learning is one of many tools we’re using to help our vehicles respond appropriately to situations on the road. We start with a rule-based system based on our own intuition as human drivers. Over time, we replace that system with deep-learned models that reveal the biggest challenges to our performance on the road. Then, we focus machine learning and deep learning tools on solving those particular problems. This approach has led us to a rate of improvement that’s unparalleled in the industry.