32(ish) Questions with an MIT Robotics Researcher (and Actor!)

[00:00:00]Hey, Alex!

[00:00:01]Hey! How's it going? Pretty good.

[00:00:03]Can you tell me a little bit about what you're working on?

[00:00:06]Yeah. So, I am a fourth year Ph.D.

[00:00:08]student at the MIT Aero-Astro Department working on human robot collaboration.

[00:00:13]so if I understand this correctly, you're in the Aero-Astro department.

[00:00:16]But your work is all about robots and people.

[00:00:19]How does that work?

[00:00:21]Yeah.

[00:00:21]So, people are surprised when they hear that Aero Astro works on robots that aren't necessarily like rockets or aircraft, right?

[00:00:27]But a lot of the same fundamental methods like controls or planning come from things like aircraft.

[00:00:33]So even though we don't work on those techniques, those same base ideas can be very useful for robots like this.

[00:00:38]Okay, so you did your undergrad, your masters, and now your PhD all at MIT - so what’s kept you here?

[00:00:44]Yeah.

[00:00:44]So it's going to sound cheesy, but, actually really the people. Right.

[00:00:47]There's a real energy here that's hard to find other places.

[00:00:50]Right? And that, works on research, right?

[00:00:52]We all work very hard.

[00:00:53]But there are people in my lab that our, world class glassblowers really good cellists.

[00:00:59]I worked on an annual project to, make a two story haunted escape room out of, plywood and two by four there’s just energy Here are people who just want to do cool, weird things, and that's really infectious.

[00:01:09]Okay, so I want to know what's going on here because I understand your work is about robots acting in close proximity to people But why is being physically near someone such a problem in robotics?

[00:01:20]So that actually brings us to a project that we have over here.

[00:01:24]So, there's a couple of reasons.

[00:01:26]First is that, with people, safety is very important, right?

[00:01:30]If you imagine a robot, say, unloading a dishwasher.

[00:01:33]Right.

[00:01:33]If a robot drops a plate or bumps over something, you know, no harm, no foul.

[00:01:38]But with people, if you bump into them, you can actually harm them or lose trust.

[00:01:42]So it's very important to make sure you have very safe systems.

[00:01:45]But secondly, people are very ad hoc right?

[00:01:47]They're hard to predict.

[00:01:48]So it's really hard to make sure you have a system that is both safe and can interact with people in that sort of way.

[00:01:55]So it's actually very applicable to this demo.

[00:02:01]Yeah.

[00:02:01]So this demo was designed by another member of our lab.

[00:02:04]That's a novel way for a robot to plan paths between two points.

[00:02:09]So here, that's these two blocks.

[00:02:11]And what's special about it is that it's able to, update its knowledge and avoid very dynamic obstacles in its environment.

[00:02:17]So if I move this block here, it's going to know to get out of the way very quickly.

[00:02:22]And the special part about this demo is that it can update its knowledge about the space and update its plan, at a thousand times per second, which means that even if people move very quickly through the space, it can always know exactly what they're doing and be able to get out of the way and act safely.

[00:02:37]So that sort of is the sort of safety that we're looking for in human robot interaction, one that can react very quickly to how people, move and interact in the same space as a robot.

[00:02:46]OK so I see here there’s a quadruped from Boston Dynamics.

[00:02:50]You talked about how these robots can do one task.

[00:02:52]Really, really well. We've seen a lot of videos of this.

[00:02:55]But then why aren’t they ubiquitous in the world?

[00:02:57]Yeah.

[00:02:58]So that comes down to the fluidness of human-robot interactions and how robots are able to understand what humans are doing and work well around them. Right.

[00:03:05]It's difficult for a robot to, really, work safely and well around a person if it doesn't really have that sort of same mental model.

[00:03:14]There's been a lot of research recently about using language models to close that gap, but there's still a lot of work to be done.

[00:03:19]Okay, so then how would you measure when someone really trusts a robot appropriately?

[00:03:24]Yeah.

[00:03:24]So trust is really hard thing to measure historically I think in the human-robot interaction literature there's been like 30 different definitions of trust that's been used.

[00:03:32]But really it's important because ultimately if a human doesn't trust a robot, they're not gonna want to use it.

[00:03:37]So there's a lot of instances we see right now where a robot gets bought by someone and then gathers dust in the corner.

[00:03:41]So understanding how to engage in correct amount of trust, can really help us make robots be more useful and more engaging with people.

[00:03:50]Okay. Sounds like you've got your work cut out for you, but when you're not in the lab, how do you like to spend your free time?

[00:03:56]Yeah. How about I show you? Let's meet over there.

[00:04:00]So here we are at the MIT Kresge Theater.

[00:04:03]So actually, while I've been in grad school.

[00:04:04]I've gotten very involved in the various theater groups around campus.

[00:04:07]It's very common at MIT.

[00:04:10]Even though we all mostly study STEM to get into various other activities in political science and the arts, and I've really gotten into this specifically.

[00:04:17]And it's been a very good opportunity for me to get involved in a new community, a new sort of creative outlet... like Shakespeare.

[00:04:30]“All the world’s a stage... and all the men and women, merely players” Sorry, I got a little carried away there.

[00:04:37]Bravo.

[00:04:38]So how did you get into acting in the first place?

[00:04:42]Yeah.

[00:04:42]So MIT requires undergrads to take classes outside their major; classes in humanities, arts and social sciences.

[00:04:48]And I took my first acting class basically on a dare, like a thing that I would never really do.

[00:04:52]But when I started, I actually got really into it.

[00:04:54]So I know you are primarily a scientist, but has anything that happened in these walls influenced your research?

[00:05:01]So a big part of research is communication, right?

[00:05:04]For things like presentations, giving talks and a lot of the same ideas and, skills that you learn in acting really translates the same idea of communication, right?

[00:05:12]In a scene that might be an emotion or a character beat, and in research it's a research paper or something, but those same ideas really come together very well.

[00:05:20]Okay. Very good advice.

[00:05:21]Now, do you have any favorite productions?

[00:05:23]I am a Lion King person myself.

[00:05:25]Good choice.

[00:05:26]I really enjoyed our production of Julius Caesar, where I played Mark Anthony.

[00:05:29]There's a lot of good monologues there, right?

[00:05:31]Friends, Romans, Countrymen, all that good stuff.

[00:05:34]Now, I do want to see a little bit more of your research, so can we head back to the lab?

[00:05:38]Yeah. Let's go.

[00:05:42]Okay.

[00:05:43]Now I'm hearing a lot about this term “Physical AI” What does that mean to you?

[00:05:49]yeah.

[00:05:49]So Physical AI, I think is sort of a question about how robots can function with more and more data.

[00:05:54]So think about like ChatGPT something that has a lot of compute, a lot of data that you put into it, trying to transition that into something like, like a robot.

[00:06:03]Right.

[00:06:04]So how can a robot learn general tasks from a lot of data?

[00:06:08]And it's definitely a field that's very interesting, very much growing right now.

[00:06:11]But I think there's still a lot of things that aren't sort of in that area that can still be very useful.

[00:06:16]So take us back to the beginning.

[00:06:18]What's the first robotics project you ever worked on?

[00:06:21]So the first Robotics project that I actually worked on in Research World, it's actually still in this area of human-robot interaction.

[00:06:27]And it was on this idea of task allocation between humans and robots.

[00:06:32]So trying to learn over, repeated tasks how robots and humans, can sort of learn who's better at what so we can allocate the right tasks to them so that the team can do the best overall as a unit.

[00:06:46]Now I want to show you the demo.

[00:06:47]Let me go real quick.

[00:06:56]Hi. Welcome back.

[00:06:57]Thanks for having us. Of course. Okay.

[00:06:59]So before we get into this demo, can you tell me a little bit about your gloves?

[00:07:04]Yeah.

[00:07:04]So these gloves are for the robot to know where I am in space.

[00:07:08]So if you might see those cameras all around the top of this lab space, and those send out a small pulse of IR light, and those bounce off these little shiny balls here and are picked up by those cameras and knows where these are in space and that a set up like this, a configuration of balls like this means, for example, left hand.

[00:07:29]And that's how we get the whole process to work with the robot.

[00:07:31]knowing where I am in space, learning how I move and be able to react to me.

[00:07:36]OK, now this demo: my understanding is that you have a painting task a person and a robot arm applying paint.

[00:07:44]but, can you elaborate on what’s really going on here?

[00:07:48]So when we designed systems like I did, you want to test them out with people in the real world.

[00:07:53]And to do that, you want to create some representative task that might be informative for how people might work in the real world. Right.

[00:07:59]So this is sort of a small little task that is representative of industrial assembly tasks.

[00:08:05]So here we have five subtasks in this painting task.

[00:08:08]That is two that will always be applied by the robot, two will always be done by me, And one may be switched between the robot or myself.

[00:08:17]So the robot's job is to mix two paints.

[00:08:20]We're not using real pain because that'll be messy.

[00:08:22]into the central bowl.

[00:08:23]So go to one tray, gets a paint, put in the bowl, other tray back to the bowl.

[00:08:28]In the meantime, my job is to secure this piece of wood that will have the paint applied to it, to this vise.

[00:08:35]So in the meantime, I'm going to be prepping the vise, unscrewing it, securing the piece, screwing it back in, And then once all that is done it is either my job or the robot's job to take a brush, get some paint onto it and apply that to the piece.

[00:08:49]So again, this is sort of a simplified model task, but is intended to model real world tasks in the small way.

[00:08:55]Okay. And how does AI fit into all of this?

[00:08:58]So there's sort of a couple components here going on.

[00:09:00]There's a sort of more traditional way of thinking about task level planning, of who will do what when, how the robot order its tasks for the, the pieces, how, understand how the robot might move through space.

[00:09:13]And there's sort of multiple techniques that I use in that whole process, like the robot’s motion uses a diffusion policy, which is similar to a lot of the AI stuff that you might see in like image generation and the sort of modern sort of AI landscape.

[00:09:26]Okay, so should we talk a little bit more about how the system actually works?

[00:09:30]Yeah.

[00:09:31]So to do that, let's actually start off the system.

[00:09:37]okay I’ll have to be here.

[00:09:42]So in the second round it's going to know that I'm going to use my left hand to secure that piece of wood down.

[00:09:47]So it's going to choose to do that task first.

[00:09:51]Which means that when I come in here, secure this piece, it's not going to get in my way of this left hand like it did last time.

[00:09:59]Now it's going to, ask me to brush the piece and that will be done the second round and learn how fast I do that task.

[00:10:09]now, in situations where you don't have much data or computing power, what does your approach let you do that other methods can’t?

[00:10:15]Yeah.

[00:10:16]So I really like to find ways to blend more traditional methods like here where I do the scheduling part of it with more modern techniques like diffusion.

[00:10:25]And I think when you have, both at the same time, you are allowed to have more capability, than if you just try to do everything with all data all the time.

[00:10:36]I’m very intrigued by your experiments that you did here.

[00:10:39]I know you tested this with 32 people, also love the number, and you deliberately nudge them to behave differently.

[00:10:45]Some using your left hand, some with the right, why engineer that much variety in, rather than just letting people act naturally?

[00:10:52]Yeah.

[00:10:52]Because in the real world, people will act differently.

[00:10:55]Like, think about a manufacturing task.

[00:10:57]People will do those jobs for like 30 years and are very ingrained in the way that they do things.

[00:11:01]However, when I bring people in for a study for this robot, they've never done this task before.

[00:11:06]So it's important for me to be able to ingrain, sorta differences in how people do things, like you see in the real world in this more constructed scenario.

[00:11:15]So your results showed that the robot getting out of someone's way mattered for how people felt about working together, not just actually how fast the job got done.

[00:11:24]So why’s that side matter as much as the numbers?

[00:11:27]That's a great question.

[00:11:28]How bout I answer that at my other lab at MIT CSAIL?.

[00:11:39]Okay.

[00:11:39]You just came from Aero-Astro, now you’re in the iconic MIT CSAIL How does that work?

[00:11:45]Yeah.

[00:11:45]So some of my lab is in Aero-Astro, and some other people are in CS and that can really happen because CSAIL, the Computer Science and AI Lab sort of overarches among the different departments to find people who are interested in similar things and can bring them together in a way that can’t happen at all universities.

[00:12:01]Okay. Couple rapidfire questions: What is your favorite spot on this campus?

[00:12:06]Probably Killian Court, which is in front of the big MIT dome.

[00:12:08]It's a really beautiful place in summer.

[00:12:10]has, grass, cool architecture, really just a nice place in general.

[00:12:14]What are your favorite actiuvities on campus?

[00:12:17]Probably to go see shows from the various groups Theater, but also there's a lot of dance groups here that have really good shows every semester.

[00:12:23]What’s your favorite Large Language Model?

[00:12:28]I use Gemini, but, you know, whatever one is, seems best at the time.

[00:12:32]OK now, what’s the name of your favorite restaurant in Boston?

[00:12:36]There's a place called Andala Coffee House in Central Square that has just the best Middle Eastern food I've ever had.

[00:12:42]Favorite actor, book, movie, GO!

[00:12:45]Okay. Actor.

[00:12:46]there’s a British actor called Andrew Scott, that has been incredible in things like Fleabag.

[00:12:51]Really good stage actor and screen actor.

[00:12:53]Just hard thing to do.

[00:12:55]Book?

[00:12:56]I'm a big Tolkein fan, so Probably Lord Of The Rings, and movie?

[00:12:59]Hard to say There's a movie called ‘In the Mood For Love’ It's a Chinese film, from the 90s.

[00:13:04]Incredible.

[00:13:05]Okay, last one: are there any unsung perks of being a graduate student?

[00:13:09]Yeah.

[00:13:10]So a big part of grad school is going to conferences, and those can happen around the world, right?

[00:13:14]So if I go to a conference, I've been to places in the Netherlands in Austria and Japan all over the place You get to travel basically for free, because it's paid for by MIT and the university.

[00:13:26]to go see these awesome places.

[00:13:28]Now, what is your dream vision for your work?

[00:13:31]What does it actually feel like to have a robot co-worker who’s working next to you?

[00:13:36]Yeah.

[00:13:36]So I think a really big idea in Robotics that I would want to follow is that robots should exist to help people, right?

[00:13:42]But if you just focus on capabilities on their own, it's hard to make sure that you're really focused on the sort of user and the end person, using the robots.

[00:13:51]So having to make sure that it really helps people in the real world is where I want it to go.

[00:13:56]Okay Alex, last question: What advice would you give to viewers, especially those who might want to get into a career like yours?

[00:14:04]Yeah, it's really just to, have your own perspective on things you see in the world, right?

[00:14:11]We live in a world right now, especially in robotics and AI of hype everywhere, right?

[00:14:15]But to sort of think outside a little bit of what that hype might tell you and what might be other angles into an area like Robotics, like human-robot interaction, that sort of comes at things that sort of comes at things from sort of the side rather than the hype straight on.

[00:14:27]All right, Rachel, it was great to talk to you I had a great time answering these questions, and I'll see you next time.

[00:14:32]Thanks Alex!