Dr Gill Pratt, head of Toyota Research Institute, talks exclusively to Megatrends about the role of AI in the auto industry of tomorrow. By Freddie Holmes
Depending on who you talk to, the concept of artificial intelligence (AI) conjures up varying visions of the future – one of a world where robots live in perfect harmony with humans, another where robotic intelligence has progressed too rapidly for mankind to control. Science fiction would suggest we still have some time to wait, but the fact is that the development of robotics and AI has reached a point at which such visions now merit serious discussion.
In November 2015, Toyota announced it would invest US$1bn to develop a standalone company that would focus on the development of AI and robotics. Shortly before the announcement, the Japanese OEM had also said it would invest US$50m in a partnership with the Massachusetts Institute of Technology (MIT) and Stanford University for the same reason.
Dubbed TRI (the Toyota Research Institute), the Silicon Valley-headquartered company is a sizeable supplement to Toyota’s efforts in driverless cars. The R&D powerhouse is supported by a second facility located near MIT. Plans to develop a third facility in Ann Arbor, Michigan are also under way, and will be a stone’s throw away from Mcity, a 32-acre autonomous vehicle proving ground. In August 2016, TRI invested US$22m into the University of Michigan’s efforts in AI, robotics and autonomous driving.
Through deep learning, robots can classify different objects roughly as well as a human being. They still don’t really understand what things mean
The primary focus for TRI is to accelerate the development of technology that can operate a car more effectively than a human driver – hardly surprising given that its parent company is the largest vehicle manufacturer in the world. A particular focus is also placed on allowing those that cannot drive the opportunity to make use of a car. Outside of driverless cars, TRI will also investigate ways of improving total mobility, moving people “across the room, across town and across the country.”
Eventually, autonomous cars will drive better than people in most circumstances
Assuming the role of Chief Executive at TRI is Dr Gill Pratt, ex-Director of the DARPA Robotics Challenge, which arguably catalysed the development of autonomous cars. Pratt’s interest in robotics dates back to his childhood where he was ‘enthralled’ with robots, and in 1979 he joined MIT. After 21 years, he joined Franklin W. Olin College to work on its automation and robotics programme, before moving on to the Defence Advanced Research Projects Agency (DARPA). He was then headhunted by Toyota.
Speaking exclusively to Megatrends, Pratt explains how robotics and AI are likely to become commonplace not only within the automotive industry, but also across many aspects of modern life.
AI makes smart cars smarter
Alongside AI, industry stakeholders may also be familiar with the terms machine learning (ML) and deep learning (DL).
AI describes anything that acts in an intelligent manner that is not controlled by a natural system, i.e. an animal or a human. ML is a form of AI that can adapt and learn from experiences. DL is a refined form of ML that uses deep neural networks – layers of algorithms that are designed to recognise patterns.
The question is, how can these technologies help the automotive industry in its pursuit of cars that can drive themselves? Can AI both drive and think like a human?
“Eventually, autonomous cars will drive better than people in most circumstances,” says Pratt, “but I don’t think there’s a person on earth right now that knows when that will happen.” AI, he continues, is an absolutely vital component to the fully autonomous car, but it needs further refinement. “Through DL, robots can classify different objects roughly as well as a human being can. They can understand the difference between a book and a train, for example, but turning that perception into cognition is still very hard to do,” he admits. “They still don’t really understand what things mean.”
For example, if a parent is holding a child’s hand on the pavement, the chance that the child is about to run across the street is different to a teenager riding a bicycle. As humans, we have the ability to draw realistic conclusions from certain situations, but robots do not share the same ability just yet.
“We have all kinds of information that we bring to bear when we see one of those scenes,” says Pratt. “There is this incredible human ability to generalise and think: ‘Well, if I were that parent or child, how would I act?’ Presently, AI systems – even DL ones – are very good at perception, but not at cognition. I think we still have a long way to go with this.”
If autonomy is to be trusted, it must drive in a substantially safer way than humans, he affirms.
Rewriting unwritten rules
In the 1950s, Isaac Asimov’s series of science fiction novels I, Robot (which were later adapted into the film under the same name in 2004) defines three laws of robotics.
Firstly, a robot must not injure a human or allow them to come to harm. A robot must also obey orders from human beings, except where such orders would conflict with the first law. In addition, a robot must protect its own existence – again, as long this does not conflict with the first or second law.
If the driverless car can eventually think by itself, many believe it too needs to operate with a set of morals to allow for ethical decision making. What does the car do if avoiding a pedestrian involves driving into oncoming traffic? Pratt is aware of the dilemma here, but contests the validity of the term ‘morals’ in this context.
“To call it ‘morals’ is a little bit of a mistake,” he suggests. Instead, Pratt believes the driverless car will need to operate through a set of rules. “There are many unwritten rules between drivers behind the wheel. Humans negotiate by looking at people in other cars and gauging their facial expressions, but we don’t have ways for an autonomous car to do that kind of negotiation,” he observes. “We’ll need to have a published set of rules that can be used amongst different manufacturers and pedestrians.”
Are robots our friends?
A common theme has developed among robotics companies; many robots share common characteristics with natural – living and breathing – things.
Autonomous cars could be safer if they were good enough, but autonomous driving where the car drives by itself is not the only way to get there
Honda’s robotics subsidiary, Asimo, has been developing ‘the world’s most advanced humanoid robot’ since 2000. Google-owned Boston Dynamics is gunning for advanced ‘human simulation’, and has developed a series of robots to assist the US armed forces. Videos posted online depict the robotic ‘Big Dog’ bounding around the woods surrounding its development centre, and the near six-foot tall ‘Atlas humanoid’ strolling alongside an engineer.
Pratt explains that there are “emotional reasons” why robots are commonly designed to mimic natural beings. “For the programmers of these robots,” he says, “it’s easier to control them if they resemble animals or people. We’ve built the world to fit ourselves, so building a robot that looks like us makes it a natural fit.”
If robots are being developed to appear familiar, should the interaction also feel familiar? Pratt thinks so.
Humans negotiate by looking at people in other cars and gauging their facial expressions, but we don’t have ways for an autonomous car to do that kind of negotiation
“There’s a very well-known effect in psychology and computer science where people will anthropomorphise a machine,” he observes. “We’ve seen this in the car business for a very long time; people refer to their car in loving terms, they name it and imagine that it has moods. The robot itself is going to very much be of that style.”
Once we have machines in our homes helping us in a variety of ways, I think we’ll see more of that bond between people and particular machines
Recalling his time working with the DARPA robotics challenge, he notes that people “bonded tremendously with the robots,” and when the machine encountered difficulty or fell down, the audience gasped. “When it was successful – even at the simple tasks – there were incredible cheers,” he added.
If consumers eventually form an emotional bond with their robot counterparts, be it their car or an automated home droid, Pratt believes it is “very likely” that AI will become a necessity in the same way that the smartphone has. “I think that will be good,” he affirms. “There are many better things to do in life than spend all your time doing dull dangerous tasks. It would be neat for us to use the technology we have on earth to do things that are more worthwhile.”
In addition, AI could have a meaningful impact at the dealership. Cars are often an emotional purchase for many drivers, with driving performance and exterior styling the primary attraction. Over time, the interest has shifted slightly toward a desire for premium qualities inside the vehicle – luxury materials, driver assistance technologies and connectivity services. There have been suggestions that the car will become less of a personal device and more of just a means of getting from A to B, but Pratt believes that a robotic personality could help to ensure the car does not simply become a domestic appliance.
“With the rise of the smart phone and the Internet, there is a tendency for young people to express their personality through information that is online, rather than in a piece of hardware sitting in the driveway,” he observes. “Once we have machines in our homes that are helping us in a variety of ways, I think we’ll see more of that bond between people and particular machines.” This could be welcome news for dealerships looking to maintain new car sales in a world where mobility as a service is increasingly popular, he suggests.
The trillion-mile problem
The conversation around autonomous cars has grown from a hushed discussion to a collective roar, but it is worth pointing out that the underlying technology has been around for a long time. For example, an autonomous vehicle developed by Stanford University and Volkswagen won the DARPA Grand Challenge more than a decade ago.
Since then, various OEMs have released timelines that outline when to expect an autonomous vehicle from their brand, and government investments around the world are providing a welcome boost to R&D initiatives. Why, then, is a future involving automated driving still up for debate? The short answer is that autonomous cars need to prove their worth, and that takes testing. And plenty of it.
“We tend to call this the trillion mile problem,” says Pratt. Globally, Toyota produces around ten million vehicles a year (10.08 million in 2015), which last around ten years, each travelling roughly 10,000 miles each year, he explains. “Multiply all that out and together, they travel a trillion miles a year. The most advanced of the autonomous car software makers – Google – has tested close to two million miles since its driverless car programme started in 2009. There’s a big difference between two million and one trillion.” He explains that this makes it very difficult to build an autonomous car that is error-free in all of those miles, and consumers will not tolerate any mistakes.
Pratt addresses the suggestion that autonomous cars are ultimately an expensive and time consuming answer to an issue that could ultimately be solved by wearing a seatbelt and not driving when drunk or distracted. He agrees, but counters the argument, suggesting that “some kind of technology that can make an evasive manoeuvre to avoid a crash” is also a solid fix.
In the same way that we have quality control in elevators – in their construction, approval and inspections – we need to have very good quality control in autonomous cars to make sure they truly are safe
As such, TRI is currently developing two different approaches to autonomous driving – dubbed ‘chauffeur’ and ‘guardian’ mode. The former can collect a consumer and take him or her home. They could even fall asleep, suggests Pratt. Guardian mode effectively observes the human driver and is ready to step in if required. This, says Pratt, is not what most people would call an autonomous car.
“It’s rather a car that has a tremendous number of safety features to prevent it – no matter what – from causing an accident,” he explains. Autonomous cars can be safer than human drivers if the technology is developed enough, but a car that drives by itself is not the only way to improve road safety, he affirms.
An existential issue
Early in 2015, a paper published by the Association for the Advancement of Artificial Intelligence, titled ‘Research Priorities for Robust and Beneficial Artificial Intelligence,’ investigated the future impact that AI could have on human life. It notes that the growing capabilities of AI “are leading to an increased potential for impact on human society,” and that “it is the duty of AI researchers to ensure that the future impact is beneficial.”
This would suggest that AI could potentially have negative, if not unsafe implications. An open letter signed by more than 8,000 experts in the field – including Professor Stephen Hawking and Tesla Chief Executive, Elon Musk – called for researchers to not create something that cannot be controlled. In 2014, Professor Hawking had warned that success in creating AI could not only be the ‘biggest’ event in human history, but also the last. “The development of full AI could spell the end of the human race,” he told the BBC.
Pratt agrees that as with any technology, there are associated benefits and dangers. Luckily for the automotive industry, he does not believe AI presents a danger. Instead, he believes the risk is in quality control. “In the same way that we have quality control in elevators – in their construction, approval and inspections – we need to have very good quality control in autonomous cars to make sure they truly are safe,” he says. “I don’t think that the ethical, scary AI issues are something we need to worry about beyond this mind-set of quality control.”
At the 2016 Consumer Electronics Show, (CES), Pratt remarked that most of the easy work had been done in autonomous driving, but the hard work is still waiting to be done. AI will be an important aid here, he says, but assures Megatrends that the automotive industry, its stakeholders and the public can sleep easy: the world is not about to be overtaken by out-of-control robot cars.
This article appeared in the Q3 2016 issue of Automotive Megatrends Magazine. Follow this link to download the full issue.