Unsupervised Feature Learning for Manipulation with Contrastive Domain\n Randomization

Robotic tasks such as manipulation with visual inputs require image features\nthat capture the physical properties of the scene, e.g., the position and\nconfiguration of objects. Recently, it has been suggested to learn such\nfeatures in an unsupervised manner from simulated, self-supervised, robot\ninteraction; the idea being that high-level physical properties are well\ncaptured by modern physical simulators, and their representation from visual\ninputs may transfer well to the real world. In particular, learning methods\nbased on noise contrastive estimation have shown promising results. To\nrobustify the simulation-to-real transfer, domain randomization (DR) was\nsuggested for learning features that are invariant to irrelevant visual\nproperties such as textures or lighting. In this work, however, we show that a\nnaive application of DR to unsupervised learning based on contrastive\nestimation does not promote invariance, as the loss function maximizes mutual\ninformation between the features and both the relevant and irrelevant visual\nproperties. We propose a simple modification of the contrastive loss to fix\nthis, exploiting the fact that we can control the simulated randomization of\nvisual properties. Our approach learns physical features that are significantly\nmore robust to visual domain variation, as we demonstrate using both rigid and\nnon-rigid objects.\n