The Kinova Innovator Spotlight series is a Q&A discussion with roboticists from all horizons sharing Kinova values of humanity, excellence and creativity. It will be a great opportunity to learn more about them, their academic path and their past, current and future work.
As our first Innovator Spotlight, we are proud to feature Anca Dragan, an Assistant Professor at the University of California — Berkeley and a Research Scientist at Waymo. She directs the InterACT Lab.
Kinova: Thanks Anca for accepting to be our first guest to our new Innovator Spotlight series.
Anca Dragan: Yes, thanks a lot for including me in this, it’s great!
K: If we go back a bit and talk about your student years. When did robotics start to be an option for you?
A.D.: Not until graduate school. I foolishly went to the Robotics Institute at CMU for my Ph.D. without any robotics experience. I was working in AI more broadly, and Illah Nourbakhsh was my AI professor and encouraged me to apply not just to the Computer Science department, but to the RI as well. I was planning on working more on the scheduling/optimization side and not so much on robots, but then CMU has this "marriage" process where you spend a month talking to faculty and figuring out who you want your advisor to be. I stumbled upon a manipulation research lab (led by Sidd Srinivasa), and I got excited not so much by the manipulation application, but intellectually by the algorithms driving the robot's motion planning. So, while I worked in computer science broadly from a very early age (programming from 4th grade, algorithms from 6th grade) [...], going into robotics, in particular, happened a bit by chance.
We tend to imagine robots and autonomous cars doing their own things magically... but actually they are doing all this in a world that contains US, the humans. - Anca Dragan, WIRED25
K.: In your 2019 WIRED talk, you talked a lot of the human-robot-interaction (HRI), why and when did it emerge as your preferred field of research?
A.D.: We would have robots generate their actions for manipulation, navigation, etc. all autonomously (at least that is where the research was heading). We'd given them their goal or objective, their constraints, and they would figure out what to do. Yet, when it came to getting robots to coordinate and collaborate with people, we'd have to design heuristics or strategies by hand. In my lab, Maya Cakmak was visiting to fix our robot's broken handovers, and she came up with this brilliant idea of adding spatio-temporal contrast. [...] But over in manipulation land, for instance, I wouldn't have to come up with a strategy for which grasp the robot should choose depending on what clutter is around the target object -- it'd figure that out by itself!
I got very excited at the thought that maybe the robots could figure out strategies for interaction by themselves as well, so that we don't have to program them in.
It was an opportunity to help bring this rational decision-making or optimal control lens to a different area, and see how far we can go.
K.: What would be a good example of HRI and how you implemented it with your Kinova Gen2 robotic arm?
A.D.: We work a lot in my lab on how you can customize what the robot does to the end-user preferences. It turns out people give you a lot of feedback and information about what they want implicitly and explicitly. So the goal is to use that to adapt to what the robot is doing.
With our Gen2 (a.k.a. Jaco), as we see in the paper “Learning Robot Objectives from Physical Human Interaction”, instead of solely moving the robot around to teach it different positions, the user applies an external torque to a robot that is already doing something. This means that rather than fighting against the force of a human, you actually use this torque to optimize the task that you are trying to do with the robot.
Above: GIF animation showing impedance control
K.: What is next for you in your lab in terms of robotics research?
A.D.: I should soon start to work with a neurologist of UC San Francisco that implanted an electrode array in the motor cortex of a tetraplegic patient. That patient wants to be able to control a PC by himself, but also feed himself.
We hope to use the Jaco arm. We will use Reinforcement Learning to take the user input and learn how to map that to perform the task better.
[...] At the same time, we will aslo experiment with techniques that try to decode the person’s intention as they are operating the robot, which is scientifically and practically really interesting. In the end, the robot should not only blindly execute, but understand what the user really wants to do.
Also, another thing that I’m excited to work on, in addition to the interaction Robot/End-user, is the interaction between the robot and the designer of the robot. The programmer designs the robot in a way “he thinks” is good.[...] But the programmers are not magic oracles, they make mistakes!. So I think artificial intelligence can be a great tool to work on the interaction between the robot and the designer of the robot. The designer would then be presented with some hypothetical scenarios that he may have not thought at first. As designers, we have to stop lying to ourselves that we can program these robots to be robust and ready to do the right thing on the first try…
K.: And finally, what challenges have you faced being a woman in a male-dominated field?
A.D.: It was easier than it was for my female mentors, and hopefully harder than it will be for my students. I've been very lucky in my career so far, but it's the implicit bias that is annoying. For instance, I come from a math Olympiad background, yet I was told to add more math to my slides in presentations because people might just assume I'm not as mathematically-savvy because of my gender. Add to this that I work on interaction with humans, which has the bias of a "soft science". These issues would really bother me in grad school, but I hope that by doing good algorithmic work and raising the next generation of strong female students we can change those perceptions a bit.
Anca first used a Kinova Mico 6 DOF arm while doing her Ph.D. in Robotics at Carnegie Mellon University with Siddhartha Srinivasa. At UC Berkeley, she is the owner of the first Gen2 Spherical 7 DOF that Kinova produced in September 2017.
To learn more about Kinova human approach of robotics, if you want to be part of our next edition of Innovator Spotlight or even share a Publication featuring Kinova, do not hesitate to contact us!
- Photograph of Anca Dragan in top hero banner: © Noah Berger
- Top hero banner source image: bongkarn thanyakij via Pexels, https://www.pexels.com/photo/red-notebook-and-smartphone-beside-computer-laptop-3774023/