Abstract
We address the problem of spatial segmentation of a 2D object in the contextof a robotic system for painting, where an optimal segmentation depends on boththe appearance of the object and the size of each segment. Since each segmentmust take into account appearance features at several scales, we take ahierarchical grammar-based parsing approach to decompose the object into 2Dsegments for painting. Since there are many ways to segment an object thesolution space is extremely large and it is very challenging to utilize anexploration based optimization approach like reinforcement learning. Instead,we pose the segmentation problem as an imitation learning problem by using asegmentation algorithm in the place of an expert, that has access to a smalldataset with known foreground-background segmentations. During the imitationlearning process, we learn to imitate the oracle (segmentation algorithm) usingonly the image of the object, without the use of the knownforeground-background segmentations. We introduce a novel deterministic policygradient update, DRAG, in the form of a deterministic actor-critic variant ofAggreVaTeD, to train our neural network based object parser. We will also showthat our approach can be seen as extending DDPG to the Imitation Learningscenario. Training our neural parser to imitate the oracle via DRAG allow ourneural parser to outperform several existing imitation learning approaches.