If you have unlimited budget you can get perfect precision robots - but that’s boring. I firmly believe robotics should be accessible to more people. That’s why I like working with robots that are cheap and usually a bit wonky. In my research I use machine learning to compensate for suboptimal control in robots.
Simulation-to-Real Transfer Learning
Whenever you want to teach a robot a new skill you either spend meticulous hours training the physical robot, resetting, and refining the action. Or if you prefer to have the robot explore autonomously you may risk the robot damaging itself or getting stuck. Therefore it’s usually better to start training in simulation. There are however problems because the actions learned in simulation don’t translate well to real environments, because of noise, friction, and because simulations generally aren’t perfectly accurate. One of my interest is in making this simulation to real transition easier.
Deep Reinforcement Learning /w Language
I’m in absolute awe of the fact that you can have computers learn to play Atari games (and now Starcraft, DOTA, and others) all by themselves. However many algorithms in this field need significantly longer to pick up a new skill than any human ever would because they lack common concepts. I’d like to imbue computers with language understanding and other “basic” cognitive skills in order to make them learn faster.
Write me to email@example.com or see the links below.