Machine Learning and Artificial Intelligence for Robotics

When Professor Brenna Argall talks about the Machine Learning and Artificial Intelligence for Robotics course she teaches in Northwestern Engineering's Master of Science in Robotics (MSR) program, she doesn't mince words when describing its difficulty. 

In 10 weeks, she says, "We cover a lot of material."

That summation is not meant to scare students, according to Argall, but instead to illustrate the significance of the coursework and how important it is for robot operation and robotic research. She recently took the time to talk more about the class and how students will be able to take lessons they learn and implement them later on in their careers. 

How do you like to describe the content of this course?

What I try to do is pull out the tools and algorithms from AI and machine learning that are used most in robots. I want to give people a little bit of background about these since it's not a requirement to take a previous course in AI or machine learning. Most importantly, I try to get students to understand why and how to use these tools within robotics.

What is your primary goal for the students in the course?

I want them to have a fundamental understanding of these tools and how they can use them. I want them to see these tools in action and how they can differ from textbook examples. Robotics exist in a three-dimensional world, and a lot of AI work depends on a discreet world where they may only be a small number of spaces to visit. AI and machine learning have a lot to do with motion control, and because of that, you need to understand how to control a robot and operate it in a safe manner. 

What do you enjoy most about the course?

It's extremely challenging, and I like that. I am proud to see when students rise to the challenge and get something working, even if it took a lot of time and effort. This course is very implementation based, and so the majority of students' grades are largely project-based. Seeing their satisfaction with their projects and how much they learned and internalized is exciting. 

Why is this course important to the overall MSR student experience?

MSR students are largely preparing to go out into industry after they graduate. With that in mind, one of the goals of the program is to prepare them appropriately for that experience. If students are going to go out into industry, they presumably will be working on a real hardware system, and that system will have intelligence and machine learning capabilities. We want them to have a baseline understanding of how these different components interact with one another. Even if you're working on a sliver of a robotic device, like the perception system, understanding the interplay of the systems is an important understanding to have. That is why I put so much of an emphasis on implementation-based projects. I feel that hands-on experience is the best way to internalize the learnings and characteristics necessary to succeed in the industry. 

What would you say are the biggest misconceptions about machine learning?

One misunderstanding I think is that there are different aspects of machine learning, and how you define the term depends on how you frame the problem you're trying to learn and what algorithms are appropriate. Statistical machine learning is all about extracting patterns from data to make new predictions, but reward-based learning paradigms are framed fundamentally differently. Those paradigms give you tools to evaluate scenarios you haven't seen before — they're more open-ended and the learning process takes longer.

Other misconceptions are that everything is learnable. Everything is certainly not learnable. Also, I think there is a fear of spontaneous AI and the notion that machines will learn something and do things radically different than what we want. The reality is we can barely get them to do what we want them to do! That doesn't mean they won't do things we're not expecting, but it will still be similar.

How will students who go through the course be able to incorporate the lessons they learn into the rest of their MSR experience or their robotics career?

Students all do an internship during their time in the program, and then they also spend one quarter doing a project with their PI at NU. That basically gives them one quarter with academic research and one quarter in either an industry or governmental lab. This class is one of the courses that prepare them to succeed in these opportunities.