Please tell us about yourself
Ashutosh Saxena , an assistant professor in the computer science department at Cornell University, shuttles between Cornell and Stanford University, where he is leading a team of scientists who are working on a $1 million open-source Robo Brain project—a large-scale computational system that learns from Internet resources, computer simulation, and real-life robot trials. The 29-year-old scientist did his BTech from the Indian Institute of Technology (IIT), Kanpur in 2004, and got a PhD in 2009 from Stanford. In a phone interview from the US, he spoke about his passion for robotics, his perception about intelligent robots harming humans, and the evolution of the Robo Brain. Edited excerpts:
What did you study?
Saxena, who joined the Cornell faculty in 2009, received his B.Tech. (Electrical Engineering) from the Indian Institute of Technology Kanpur in 2004 and his M.S (Computer Science) and Ph.D. in Machine Learning from Stanford University in 2006 and 2009, respectively.
How long has your team worked on the Robo Brain?
We have been building different pieces of algorithms for this project in my own lab for the last five years. We first worked on perception. Then, for three years, we worked on planning, and for the last one year, we worked on language. (Perception, planning and language understanding skills are critical for robots to perform tasks in a human environment). This was by design, and when all the pieces were in place, we launched the Robo Brain project, which formally went live in July.
When will robots start accessing the Robo Brain?
Right now, we have connected four universities—Stanford, Cornell, Berkeley and Brown. By the year-end, we will expand it to 10. And by 2015, we should be able to connect about 100 institutions. Finally, by 2016, we hope to open up this project to the public.
There are similar initiatives like the Never-Ending Language Learning (NELL) project at the Carnegie Mellon University and International Business Machines Corp.’s Watson supercomputer…
Yes. I’ve seen a lot of effort being put into these kinds of projects. NELL handles the language part, online learning systems, etc., well, and Watson deals with context learning efficiently. There are other initiatives for images, etc. We are collaborating with some scientists for computer vision, for instance, and will do so with others too going forward. Our aim is to build this huge learning infrastructure for robots (Robo Brain).
There is a fear that intelligent robots may dominate us some day, if they are not programmed to be human.
That is one thing we are careful about. We provide controlled data to the robots, and people can correct the learning process. You can visit Robobrain.me to see how the learning process works. It is crowdsourcing feedback.
How did you get interested in the an offbeat, unconventional and uncommon career such as robotics?
I was interested in robotics since childhood. I was born and studied in Bareilly (a district in Uttar Pradesh), where I used to make robots even in school. I used to tear apart devices like music players and cassette recorders so that I could use their motors to make the robots. These were very small toy robots, around 8 inches by 6 inches, which could move around and accomplish some task. My mother was very supportive (his father had passed away), and I had a robotics room to myself. Today, my brother too is in the field of medical robotics in Canada.
How did your education at IIT-Kanpur and Stanford help you in the task?
It was a very good learning environment. I got to publish many papers, and there were many robotics competitions, and I won a few, which gave me a lot of experience and I got to learn a lot about hardware. At Stanford, I got more interested in programming robots as opposed to just designing them. One of my projects in 2006 was about a robot that could take dishes and put them in a dishwasher. Another one was designing software to convert a 2D picture into a 3D one (with Make3D, you can take a two-dimensional image and create a three-dimensional “fly around” model). The project was co-led by Andrew Y. Ng. After this project, I got a job offer from Cornell. I now also spend time at Stanford for the Robo Brain project, and will be there for a few years since it’s a multi-university project and requires collaborative work.
How does your work benefit the community?
In Saxena’s Robot Learning Lab in Upson Hall, graduate students work with one-armed industrial robots and some home-grown contraptions, teaching the machines, in effect, how to learn, so they can adapt to the cluttered environments of homes, hospitals and nursing homes and the highly unstructured behavior of the humans they work with.
Ordinary industrial robots carry out precisely programmed motions: You can program one to pick up a cup from a specific location in the cupboard and place it in a specific location on the table. If the cup isn’t there, too bad; if the table isn’t there, even worse. So Saxena’s team adds a computer brain and a Microsoft Kinnect 3-D camera. The robot can be shown a series of coffee cups, notice what features they all have in common and learn to recognize the next coffee cup it sees, even if it’s slightly different, and pick it up by whatever sort of handle it has. Similarly, a robot can learn to scan an unfamiliar kitchen to find the sink, dishwasher and cupboards so it knows where and how to place the cups.
Understanding what humans are doing is harder, but the researchers have taught robots to observe humans and identify a number of common activities. The next step, Saxena says, is to link perceptions with actions so a robot could, for example, notice that a patient hasn’t taken his medicine, remind him, and bring the bottle to his bedside.
Any plans for India?
We are looking for collaborations because our software can be useful in sectors such as manufacturing. The IITs will be on our radar when we expand in a year’s time. But we have not initiated any formal talks as yet.