Using Simulation and Domain Adaptation to Improve Efficiency of Deep Robotic Grasping

In this work we extensively evaluated the
effect of using simulation and domain adaptation on vision-based robotic grasping with a total
of more than 25,000 physical test grasps on a set of 36 diverse objects unseen during
training. Our proposed approach, takes as input synthetic
images, generated by our simulator, and produces adapted images that look similar to real-world
ones. We then train a deep vision-based grasping
network with adapted and real images, which we further refine with feature-level domain
adaptation. We show that by using synthetic data and domain
adaptation, we are able to reduce the number of real-world samples needed to achieve a
given level of performance by up to 50 times, using only procedurally-generated objects
in simulation. Our method does not require paired examples
and can perform well, even in the absence of labeled real-world samples.

, , , , ,

Post navigation

Leave a Reply

Your email address will not be published. Required fields are marked *