Q&A: Cathy Wu on creating algorithms to soundly combine robots into our world | MIT Information


Cathy Wu is the Gilbert W. Winslow Assistant Professor of Civil and Environmental Engineering and a member of the MIT Institute for Information, Programs, and Society. As an undergraduate, Wu gained MIT’s hardest robotics competitors, and as a graduate pupil took the College of California at Berkeley’s first-ever course on deep reinforcement studying. Now again at MIT, she’s working to enhance the circulate of robots in Amazon warehouses underneath the Science Hub, a brand new collaboration between the tech big and the MIT Schwarzman School of Computing. Outdoors of the lab and classroom, Wu may be discovered operating, drawing, pouring lattes at residence, and watching YouTube movies on math and infrastructure by way of 3Blue1Brown and Sensible Engineering. She not too long ago took a break from all of that to speak about her work.

Q: What put you on the trail to robotics and self-driving automobiles?

A: My mother and father at all times wished a physician within the household. Nevertheless, I’m unhealthy at following directions and have become the mistaken form of physician! Impressed by my physics and pc science courses in highschool, I made a decision to check engineering. I wished to assist as many individuals as a medical physician might.

At MIT, I appeared for functions in vitality, schooling, and agriculture, however the self-driving automotive was the primary to seize me. It has but to let go! Ninety-four p.c of significant automotive crashes are brought on by human error and will probably be prevented by self-driving automobiles. Autonomous automobiles might additionally ease site visitors congestion, save vitality, and enhance mobility.

I first realized about self-driving automobiles from Seth Teller throughout his visitor lecture for the course Cellular Autonomous Programs Lab (MASLAB), during which MIT undergraduates compete to construct the most effective full-functioning robotic from scratch. Our ball-fetching bot, Putzputz, gained first place. From there, I took extra courses in machine studying, pc imaginative and prescient, and transportation, and joined Teller’s lab. I additionally competed in a number of mobility-related hackathons, together with one sponsored by Hubway, now generally known as Blue Bike.

Q: You’ve explored methods to assist people and autonomous automobiles work together extra easily. What makes this drawback so onerous?

A: Each methods are extremely complicated, and our classical modeling instruments are woefully inadequate. Integrating autonomous automobiles into our present mobility methods is a large endeavor. For instance, we don’t know whether or not autonomous automobiles will minimize vitality use by 40 p.c, or double it. We want extra highly effective instruments to chop via the uncertainty. My PhD thesis at Berkeley tried to do that. I developed scalable optimization strategies within the areas of robotic management, state estimation, and system design. These strategies might assist decision-makers anticipate future situations and design higher methods to accommodate each people and robots.

Q: How is deep reinforcement studying, combining deep and reinforcement studying algorithms, altering robotics?

A: I took John Schulman and Pieter Abbeel’s reinforcement studying class at Berkeley in 2015 shortly after Deepmind revealed their breakthrough paper in Nature. That they had skilled an agent by way of deep studying and reinforcement studying to play “Area Invaders” and a set of Atari video games at superhuman ranges. That created fairly some buzz. A yr later, I began to include reinforcement studying into issues involving combined site visitors methods, during which just some automobiles are automated. I spotted that classical management strategies couldn’t deal with the complicated nonlinear management issues I used to be formulating.

Deep RL is now mainstream however it’s in no way pervasive in robotics, which nonetheless depends closely on classical model-based management and planning strategies. Deep studying continues to be vital for processing uncooked sensor knowledge like digital camera pictures and radio waves, and reinforcement studying is regularly being included. I see site visitors methods as gigantic multi-robot methods. I’m excited for an upcoming collaboration with Utah’s Division of Transportation to use reinforcement studying to coordinate automobiles with site visitors indicators, decreasing congestion and thus carbon emissions.

Q: You have talked in regards to the MIT course, 6.007 (Alerts and Programs), and its influence on you. What about it spoke to you?

A: The mindset. That issues that look messy may be analyzed with frequent, and generally easy, instruments. Alerts are reworked by methods in numerous methods, however what do these summary phrases imply, anyway? A mechanical system can take a sign like gears turning at some velocity and remodel it right into a lever turning at one other velocity. A digital system can take binary digits and switch them into different binary digits or a string of letters or a picture. Monetary methods can take information and remodel it by way of thousands and thousands of buying and selling choices into inventory costs. Folks absorb indicators day by day via commercials, job gives, gossip, and so forth, and translate them into actions that in flip affect society and different folks. This humble class on indicators and methods linked mechanical, digital, and societal methods and confirmed me how foundational instruments can minimize via the noise.

Q: In your venture with Amazon you’re coaching warehouse robots to choose up, kind, and ship items. What are the technical challenges?

A: This venture includes assigning robots to a given process and routing them there. [Professor] Cynthia Barnhart’s crew is concentrated on process project, and mine, on path planning. Each issues are thought-about combinatorial optimization issues as a result of the answer includes a mixture of decisions. Because the variety of duties and robots will increase, the variety of doable options grows exponentially. It’s known as the curse of dimensionality. Each issues are what we name NP Exhausting; there will not be an environment friendly algorithm to resolve them. Our purpose is to plot a shortcut.

Routing a single robotic for a single process isn’t tough. It’s like utilizing Google Maps to search out the shortest path residence. It may be solved effectively with a number of algorithms, together with Dijkstra’s. However warehouses resemble small cities with tons of of robots. When site visitors jams happen, clients can’t get their packages as shortly. Our purpose is to develop algorithms that discover essentially the most environment friendly paths for the entire robots.

Q: Are there different functions?

A: Sure. The algorithms we take a look at in Amazon warehouses may someday assist to ease congestion in actual cities. Different potential functions embody controlling planes on runways, swarms of drones within the air, and even characters in video video games. These algorithms may be used for different robotic planning duties like scheduling and routing.

Q: AI is evolving quickly. The place do you hope to see the large breakthroughs coming?

A: I’d wish to see deep studying and deep RL used to resolve societal issues involving mobility, infrastructure, social media, well being care, and schooling. Deep RL now has a toehold in robotics and industrial functions like chip design, however we nonetheless should be cautious in making use of it to methods with people within the loop. In the end, we wish to design methods for folks. At present, we merely don’t have the precise instruments.

Q: What worries you most about AI taking over an increasing number of specialised duties?

A: AI has the potential for super good, however it might additionally assist to speed up the widening hole between the haves and the have-nots. Our political and regulatory methods might assist to combine AI into society and reduce job losses and earnings inequality, however I fear that they’re not outfitted but to deal with the firehose of AI.

Q: What’s the final nice e-book you learn?

A: Keep away from a Local weather Catastrophe,” by Invoice Gates. I completely beloved the best way that Gates was in a position to take an overwhelmingly complicated matter and distill it down into phrases that everybody can perceive. His optimism conjures up me to maintain pushing on functions of AI and robotics to assist keep away from a local weather catastrophe.


Leave a Reply

Your email address will not be published. Required fields are marked *