RLCapture: A deep reinforcement learning based control strategy forswitching between motion capture inspired controllers.

Making robots walk and balance as well as humans is extremely difficult. New techniques involving machine learning have shown promise in getting robots to mimic the movements of humans recorded using motion capture
technology widely used for videogames and movies. While these techniques show promise, they are still in development, and have difficulty switching between behaviours. Its still very difficult for robots to go from standing still to running. They also fall over very easily when pushed or tripped, since they don’t have a concept of reacting to pushes in the same way that people do. This research aims to solve this problem by training robots to switch between mimicked behaviours as they move, so that they can learn how to react from human examples. Ubisoft will benefit by utilizing this technology to create more realistic games where characters can react, move, and fall more like real people do.

Faculty Supervisor:

James Forbes

Student:

Kevin Bergamin

Partner:

Ubisoft Divertissement

Discipline:

Engineering - mechanical

Sector:

Information and communications technologies

University:

Program: