Learning Sensorimotor Robust Walking Controllers
Abstract
biped that can be subject to external pushes and terrain distur-
bances. Instead of the classical approach where a controller based
on character pose is hand designed and tuned manually, a senso-
rimotor approach is explored that has the potential of scaling up
to more novel tasks, characters, and environments. A pre-existing
controller is used to bootstrap the learning process and help deter-
mine the feasibility of the overall approach. Taking inspiration from
living animals that display incredibly agile motions that are learned
in an on-line setting, the long term goal is to develop techniques for
automatically learning similar controllers in a scalable fashion and
without much manual intervention.
Video
PDF
The report detailing the technique used can be found here
