Grasp2Vec: Learning Object Representations from Self-Supervised Grasping Eric Jang *,1, Coline Devin *,2,†, Vincent Vanhoucke1, and Sergey Levine1,2 *Both authors contributed equally 1Google Brain 2Department of Electrical Engineering and Computer Sciences, UC Berkeley †Work done while author was interning at Google Brain {ejang, vanhoucke, slevine}@google.com 1coline@eecs.berkeley.edu Abstract: Well structured visual representations can make robot learning faster and can improve generalization. In this paper, we study how we can acquire effec- tive object-centric representations for robotic manipulation tasks without human labeling by using autonomous robot interaction with the environment. Such rep- resentation learning methods can benefit from continuous refinement of the repre- sentation as the robot collects more experience, allowing them to scale effectively without human intervention. Our representation learning approach is based on ob- ject persistence: when a r
