Action-Conditioned Convolutional Future Regression Models for Robot Imitation Learning

摘要

Based on what is seen (i.e. visual input), humans are able to visually predict (i.e. regress) what the scene will look like after taking a certain action. Further, humans are able to take advantage of such predictions to select optimal actions for the task they are working on. Using example videos, robots can also learn to visually imagine the future consequence of taking an action. This can be viewed as learning a function mapping a raw image frame (conditioned on a particular action) to the future image frame. Once learned, the future regression function can be combined with an action policy learning framework (e.g. reinforcement or imitation learning), enabling better robot action learning for given tasks.