The world is almost ready for AI and machine learning to take over vast swaths of industry. But will AI be ready for the job? KU Leuven alumnus Pieter Abbeel is using AI to make an impact in ways that are surprisingly easy to imagine, and closer than you might think.
I’m doing the family dishes at the kitchen sink when I glance down and see the little black robot slam into my foot. The robot bounces, turns at a right angle, and heads off again, sucking up dirt all the while. I used to be happy with my little robot Hoover, watching it fulfil its purpose with every pass along my floor, a very 21st century image finally coming to pass. But I’ve been talking with Pieter Abbeel, AI wunderkind and professor of electrical engineering and computer science at the University of California – Berkeley, and he’s got me thinking about my pathetic little robot. Whereas once I was pleased, I now find myself vaguely disappointed, unable to shake this one persistent question: Why can’t my Hoover do the dishes?
And there’ve been so many dishes to do at home during the Corona crisis. Ok, I know: get a dishwasher. But if the computers inside robots can perform almost any computational task, why are our robots stuck physically performing just one task like a simple machine? To put it another way, if we gave it arms, could my Hoover learn to do the dishes? Or maybe learn to do something even more useful in the next crisis, like care for the sick?
That’s where Pieter Abbeel comes in. As a professor, as the director of the Robot Learning Lab at Berkley, and as a serial entrepreneur, Abbeel has already established an impressive academic and business record in bringing robots out into the real world. If anyone is going to teach robots to adapt to new situations, it’s him … perhaps because that’s what’s he’s done himself, at KU Leuven, at Berkeley, and now in his latest role as the unlikely voice of an AI revolution.
In high school at Sint Michielscollege in Brasschaat, Abbeel wasn’t the prototypical computer geek. He loved basketball, and while he enjoyed math and science, he enjoyed essentially every other class as well. ‘I felt literally everything was interesting,’ Abbeel says. ‘History, geography, physics, math, languages, literature … I thought it was all fascinating. But then more and more I focused on math and science to try to understand fundamentally how the world around us works and how, through that understanding, you can build new things. Like, how do you build a bridge, what are the principles?’
When he was entering KU Leuven, Abbeel had to choose between studying engineering or pure math and physics. For him, the answer was consistent with an outlook he’s maintained his whole life: ‘Ultimately in engineering, you’re going to learn a lot of math and science, but also at the same time have some emphasis on how to put it to use … for me, it made it a bit more exciting.’
Exciting or not, the KU Leuven engineering programme was demanding. ‘[It] was intense, we covered an enormous amount of ground, and we only had finals once a year in June, so I had to learn how to manage my time throughout the year.’ In addition to his school workload, Abbeel played basketball for the KU Leuven team, as well as his club team back home. Perhaps it’s no surprise that he was a point guard, tasked with seeing the whole court and determining the best way to bring his team along to reach its goal. He sees this inevitability differently, ‘I’m 182 centimetres, so as the smallest one on the team, that’s what you play. (laughs)’
After his first two years, Abbeel chose to specialise in electronics over computer science. ‘I wasn’t sure which one to choose, but ultimately it seemed like to me “Ok, that’s the foundation, you can’t build computers without the electronics, so I want to understand how this is all built, rather than just using it.”’
Abbeel’s five years at KU Leuven finished in 2000, but the next step wasn’t clear. ‘What really started fascinating me was artificial intelligence, this notion of building a network that could think … you could write a piece of code, and you’re the one programming it, but then after you write it, it can beat you to something, that it would be possible to write a piece of code that could play checkers or chess better than I could play it, even though I wrote the programme … that’s where the initial intrigue came from, building a system that could somehow do better than you could do yourself, that seemed just so fascinating.’
That’s where the initial intrigue came from, building a system that could somehow do better than you could do yourself, that seemed just so fascinating.
‘I was talking with [Professor] Bart de Moor, and he was working in system theory … he had spent his post-doc at Stanford, and the gist I got was that for AI, if you wanted to learn more, you could learn more at Stanford than anywhere else in the world. For me, that was like, one, that’s exciting, going to a place where you can learn more than anywhere else, and two, I was intrigued with spending a year in the United States. I’d never been there, not to mention California.’
Abbeel spoke with a few of his professors such as de Moor and Luc van Gool, and they told him that once he got out there, he was going to want to stay. ‘And I said, I don’t think so, I just want to go learn, I’ll be back next year, I look forward to doing a PhD in Leuven after having learned new things for a year. They said, “Yeah, we’re not counting on that.”’
In the year 2000, moving to California was an entirely different undertaking than today. The .com boom was still happening and it felt like a different world, where big companies lined the roads and the tech sector was surging. He soon saw that his professors at Leuven knew what they were talking about. ‘Sure enough, I would say after one month I was just like, ok … I started to realise the concentration of activity in artificial intelligence at Stanford, as well as a whole different level of passion with which people are applying themselves to what they’re doing. At Stanford, it felt like everyone who was there had all found a way to live their passion full time.’
Fortunately for Abbeel, ‘The education in Leuven is extremely strong. When I arrived at Stanford I realised how much I had learned … and I found that very comforting. My professors in Leuven were pretty confident about that, they said to me, “Don’t worry. Yes, it’s the best of the best from all over the world, but you’ll have been prepared having studied here.” It was a nice confirmation when I got here that indeed, my education in Leuven really prepared me fantastically for the things I was then going to learn in Stanford.’
Abbeel did indeed stick around at Stanford, beginning his PhD there under his advisor Professor Andrew Ng. Yet Abbeel learned more than simply the latest developments in AI, he also learned how to deliver that knowledge to a wider audience. ‘Two years into my PhD I asked [Professor] Ng what I should I learn, I’d already taken a lot of classes and felt like maybe I’d covered it. He said “You have the foundations, but the thing that will limit you at this point is if you don’t become equally good at communicating. That means writing, that means presenting. You should look into the writing classes at Stanford, the public speaking classes at Stanford. Not math, not engineering, communication.” It was really good advice, it helped a lot.’ After a few classes and reading up on any books he could find at the Stanford bookstore, ‘All of the sudden everything clicked, now I can do this for myself, see why the suggestions on my papers came back to me the way they did … I could see the principles now. At this point my job might be more communication than math and science. (laughs)’
Abbeel has used his newfound communication skills to improve his work as a professor, but also in unexpected ways, for instance serving online as a sort of ambassador for the robotics and AI industry, as well as securing a part in a Verizon Wireless commercial that played nationally across the United States.
After Abbeel finished his studies at Stanford, he moved across San Francisco Bay to Cal – Berkeley. In addition to teaching, he had a newfound role as director of the Cal Robot Learning Lab, which he thought might prove challenging enough for a lifetime. ‘When I started at Berkeley, the direction I was headed was robotic manipulation, how to build the intelligence, the AI software that enables robots to manipulate any object the way we humans manipulate objects. In 2008, that seemed like it would be my research agenda forever, we weren’t that far along at all, nobody was far along.’
Abbeel is one of the world leaders in reinforcement learning, and this was a chance to truly develop his work. ‘Reinforcement learning is the science of how to make machines learn from their own experimentation in the world, and gradually over time through trial and error become better at doing things. You reward [an agent] for making progress, and over time it just starts figuring out what to do to get higher and higher rewards.’
This way of thinking about reinforcement and reward naturally bleeds into life of a college professor. ‘The science of teaching is very fascinating and very close to machine learning because it’s all about how you get either a person or an AI agent to acquire a new skill, acquire a new knowledge they didn’t have before, and what’s the best way to do this. Now, with people motivation is a much bigger factor than let’s say with AI agents, where motivation may be less of a factor, because you can just force them to train, train, train. Whereas with people if they’re not excited about what you tell them, well, they might go do something else, right? And so you might have the most important thing they should be learning, but if you can’t get them excited they’re not going to pay attention.’
At the Robot Learning Lab, Abbeel was able to create a team with unparalleled focus and effectiveness. ‘They build up world-level expertise in a single domain, and a level of expertise nobody else in the world had in that domain, kept building that expertise, kept making research breakthroughs, so that at some point the path to a real-world impact was something you’d start working on.’
‘What started happening around 2013-14, we started making a lot of progress, where all of the sudden … we were able to get results that seemed impossible just a few years before that. And that’s when I started thinking about that again, it’s like, “Ok, now this whole idea of more intelligent robots actually seems like something we’re actually starting to do, it’s not just a lab thing, this might actually become a reality. Let’s think about when and where we can turn this into reality.’”
AI to IPO
Turning research breakthroughs into commercially viable products didn’t happen overnight; Abbeel’s career had been in academia, not business. Fortunately, he was in the neighbourhood of Silicon Valley, and business expertise was readily available. ‘From 2015-17 I spent two years at OpenAI in San Francisco, where Elon Musk was one of the early investors in that project … which was more pure research, not trying to put things in the real world, but make more progress. And it was then, in September 2017 that we decided that, look, the time is right, we spent so much time making robots do things that seemed impossible just recently, and now we can make them do it, let’s see what we can do to make this [have an] impact in the real world.’
But starting a company is more than just having an idea. ‘When starting a company I’d say the most important two metrics are: are you after a big market that seems you could have a path into, and two, do you have a team that is the right fit for that, that might be some of the world-leading experts.’
Abbeel began working with his grad students to develop AI to grade assignments in the STEM fields. This company, Gradescope, recently exited a successful round of fundraising. But he had his sights set on bigger goals for integrating AI into commerce.
What are the things an AI system can do in the next couple of years, where can they have the most impact?
Abbeel and three other researchers came together in 2017 to start Covariant, which went into stealth mode for the next two years, looking for an industry ripe for their AI. ‘We can build the most advanced AI robotic systems, but exactly where are they going to be useful? It’s not that we can just say, “Hey, whatever a human does a robot can do it.” No, no, no. You need to think more carefully. What are the things an AI system can do in the next couple of years, where can they have the most impact? So we went around and we met with 200 companies working in many many industries … and from that it became very clear to us, the space that was just exploding in demand was logistics. The whole fulfilment industry … goods stored in a warehouse, and then taking things out of a warehouse and actually have them shipped somewhere else was the space, that’s growing at an unprecedented pace.”
Even within a warehouse, Covariant robots aren’t going to do every job. Abbeel’s team found just the right place for the decades of specialised expertise his particular team had built up. ‘Automation had already existed for what you do with your legs as a person, conveyor belts, robots on rails, mobile robots, what did not exist was what people do with their hands, how do you automate that? Now of course the whole crisis that’s happening right now, everyone knows about being resilient, robust, and actually there’s even more demand for automation because people get sick, you still need groceries, you still need your medication, you still need everything. So you really want even more automation to be even more robust.’
The future of AI
A new wave of malleable robots is about to hit the workforce, and we’ll have to adapt our economic and social lives to this new reality. But as the Corona crisis has shown once again, we’re a pretty adaptable species, and a future with smarter robots could be a real benefit to humanity. ‘Robots don’t get the kinds of viruses humans get,’ says Abbeel. ‘So if a robot can pick and place and pack, then you know things can keep moving. The other [benefit] is more flexible manufacturing. Can we build manufacturing lines where, when you all of the sudden want to manufacture a lot of masks or other protective equipment, then you can just say, ok, let’s switch it around, now we’re making something else. That’s not easy right now; whenever things are automated they’re very rigid, they’re for the same thing over and over and over, so how can you make your robots smarter so they’re more flexible and can today make one thing and the next day make something else? That’s a big open question, but one that we’re working pretty hard on.’
This implementation of Covariant AI has already started; you can go on YouTube right now and see a Covariant robot ceaselessly sorting packages at a warehouse in Belin for an hour. Still, full implementation remains some years away. ‘The big difference between academic work and industry work is the level of reliability, the level of autonomy [you need] to achieve. We all see this with self-driving cars. Why don’t we have self-driving cars yet? Not because we can’t do a demo around the block, we’ve been able to do that for 20 years, it’s because we need a certain level of reliability, right? And that’s what’s so special about what we do at Covariant, the level of autonomy, the level of reliability consistently achieved by these robots in the face of so many different things these robots can encounter in the warehouse, it’s always different challenges coming through.’
After speaking with Abbeel, I think have a better handle on why my household robot can’t do my dishes yet. The learning curve is steep, and while certain AI and machine learning goals have already been met in warehouses, a reliable robot at home will need to break fewer dishes than I do. Probably not a realistic goal today, but it will be someday soon.
And on that day, when my robot is finally ready to do the dishes, I’ll be ready to let it.