The automotive industry has come to the conclusions that fully autonomous vehicles are not as close as some people thought three to four years ago. But that does not mean automakers and suppliers have given up on providing vehicles the ability to safely and reliably drive themselves, Nvidia Senior Director of Automotive Danny Shapiro told Automotive News Europe Managing Editor Douglas A. Bolduc.
Why have the bold predictions about putting autonomous vehicles on the road changed over the last few years?
The industry has realized that fully autonomous vehicles aren't as close as we thought they were three or four years ago. It's a really hard problem to solve. This means more computing, more sensors and more software is required.
Has the enormity of the problem caused some people to back away?
What we are seeing is people saying, "We can still bring technology to market that is the foundation for autonomous driving, but there is still going to be a human in the loop."
What will we see?
For now it is Level 2-plus, which can be a really robust system that can prevent a lot of accidents or save the driver if something bad happened like the person falls asleep or has a medical emergency. Humans still get distracted so we can prevent accidents that happen because of those human errors, even if we are not fully autonomous.
Is it more realistic to say that we won't see fully autonomous cars on the road until 2030?
If you are talking about the Holy Grail -- the car that will pick you up anywhere and take you anywhere -- yes, that is quite far out. I don't think it's that far out if you mean very specific kinds of deployments. It could be robotaxis or shuttles on fixed routes or in geofenced areas. I think those things are still moving forward with great progress. When it comes to highway auto pilots and hub-to-hub trucking, I think we are going to see a lot more in terms of the movement of goods before [the movement of] people.
Could you elaborate?
This means delivery bots of all shapes and sizes. We are involved in virtually all these delivery trial programs. It could be cars, trucks, shuttles. Those are all still developing full speed ahead. I think what people are realizing is that it's going to take a little bit longer to achieve a level of safety that we should be focused on delivering.
There has been one fatal accident involving an AV with a safety driver as well as a recent non-fatal accident. Have these events forced you to double down on what you are doing to make AVs as safe as possible?
We have always had that practice. I think Nvidia and the industry underestimated the complexity at the outset. I don't think it was a matter of us not being safe. This is something that has never been done before. We are figuring out how to solve these challenges -- and on the way we realized there is a lot more to it than was initially planned.
What is being done to address this overestimation?
It has shown the need for more high-resolution sensors on the car. The need for diversity and redundancy of these sensors, of the algorithms, of the systems to ensure the highest level of safety that we possibly can. There is this insatiable desire for more computing because it's not just one algorithm running. You have dozens of different neural nets that are all running in the car simultaneously, processing all the different data and analyzing even the same data stream. There is one neural net looking for lines in the road, one looking for pedestrians, another looking for street signs, another looking for cars and trucks. Whatever it may be. Therefore, the complexity of the software is enormous. That's why we hear the customers that we are working with such as Mercedes-Benz saying, "We need more computing horsepower in the car." That's because the complexity of the software is growing and growing.
Is creating a safe, reliable fully autonomous car the toughest challenge Nvidia has ever undertaken?
It's like comparing apples to oranges because we are also trying to help cure cancer. Is that harder? It's difficult to say. They both involve saving a lot of lives. That's what makes it really interesting for Nvidia. Working with companies to solve the world's most challenging problems is our mission. AI computing is fundamental to that. There is a lot more real-time concern over the autonomous car versus going off and solving some of these other problems that are very complex. When you are driving an autonomous car it's life or death, second by second, so you could say there is potentially more on the line there. There is incredible complexity in trying to analyze human behavior. Predict human behavior. Humans are not predictable.
What is being done to help offset some of that unpredictability?
If we were to remove other human road users and their decision-making from the equations it would be a lot easier problem to solve. Therefore, I think where we are going to see deployments in places where we can remove that human randomness. This could be dedicated roadways, fixed routes and geofenced areas. There could be cities that ban human-driven cars in certain areas or they have humans on a different plain, such as having autonomous vehicles on one level of a highway and pedestrians on another. I think there are a lot of different ways to simplify the problem. But I think creating the autonomous vehicle is one of the world's most complex problems.
What are the other takeaways from the accidents that have involved AVs?
Some of these accidents weren't caused by the same level autonomous vehicles that we are planning for. But I think the public seizes on this and it's man versus machine and the machine just scored one and the humans zero. In that case it becomes a big news story. It's amazing how many people are killed or injured every year [by cars driven by people] and the world goes on. Nobody pays attention but this one [accident involving an AV] ... .
What is being done to overcome the "fear of the unknown" connected to AVs?
There's definitely an awareness and education piece to this, which is why we joined PAVE [the partnership for autonomous vehicle education]. It is a U.S. initiative that includes Mercedes, Nvidia and a bunch of others. It was rolled out at this year's CES in Las Vegas. Autonomous vehicle education is crucial because most people don't know what is in their cars today, what the technology is capable of and what is coming in the future. They have never experienced or read a story about it and they think of the Terminator when they think of artificial intelligence. There is this fear of the unknown. But I sincerely believe once people have a chance to experience something that can almost instantly change their perception about it.
What are some of the limitations Nvidia has discovered from on-road testing of AV solutions?
There is so much on-road testing going on that involves safety drivers. When we test we have a safety driver as well as a safety engineer. They are instructed to take over if they feel uncomfortable with what the car is doing. That's good and bad. The issue is that they tend to take over prematurely. Or maybe they will take over during times when they really didn't need to, but they didn't want to risk it. That's kind of universal. Therefore, the problem is that you are never really fully testing whether the system would have avoided the collision.
What is the alternative?
This is where we turn to simulation. We are very heavy user of simulation. We're working with many customers to develop our Drive Constellation system, which does full software and hardware in-the-loop testing. We can then create all kinds of dangerous scenarios. Run them in the safety of the Cloud and not put anyone in harm's way. We can truly determine whether there an impact or not? If there was, we go back and we figure out what happened and we correct it. We can test the detection algorithms, the path planning, the mapping, the whole stack before we put it on the road. That is critical.