About Me

holding_chicken

Hi! My name is Keith Gould. I’m a software consultant, and I’m interested in making robots move beautifully. I’m primarily interested in legged movement via neural network control, though I’m still studying the fundamentals on a two wheeled balancing robot named Beaker.

I’m interested in neural network control because I think there are hard limitations to how far we can go with the classic techniques (such as differential equations for equations of motion, and hand-made models in general). For example, Boston Dynamic’s Atlas robot (which, given their papers and talks, uses classic control) jumps wrong and breaks an ankle. How quickly (if ever?) could it learn a new gait? If Boston Dynamics does indeed use classic control methods, Atlas would probably not be able to learn any sort of new gait without human help. Even without disasters, there is not any published learning capability for Atlas in-situ. I love that neural networks can always learn, always improve, always adjust.

Currently, there are not too many robots for sale that move via neural networks. This indicates how far we are from solving control with neural networks. I’d like to be part of the effort. Hopefully along the way we can make metal nimble 🙂

FAQ

  • Why Legged Robots?
    • It’s still an unsolved problem. Even the best bipeds (mentioned above) are clumsy and brittle.
    • Also, my mom is a gait analyst, so I grew up watching people walk, and find the subject really fascinating.
  • Inspiring Links
    • This post by Sridhar Mahadevan really drove home why tabula rasa reinforcement learning is not the way to go.
    • Even though I’m not a fan of using models for robotic control, I was impressed with this paper. It is what inspired me to try it out.
  •  Github?
  • Instagram?