This work was accepted to RSS 2022

Abstract

The success of deep learning depends heavily on the availability of large datasets, but in robotic manipulation there are many learning problems for which such datasets do not exist. A common approach to a lack of data is data augmentation. However, because the types of tasks and data differ, existing methods cannot be easily adapted to manipulation tasks. In this paper, we argue that augmentations should be valid, relevant, and diverse. We use these principles to formalize augmentation as an optimization problem, with the objective function derived from physics and knowledge of the manipulation domain. This method applies rigid body transformation to trajectories of geometric state and action data. We test our method in two scenarios: 1) learning a classifier of state validity for rope manipulation, and 2) learning the dynamics of pushing a set of rigid cylinders. These two scenarios have very different data and label types, yet in both scenarios networks trained with data produced by our augmentation method generalize better than those trained with no augmentation or with a naive augmentation method. We also show how our augmentation method can be used for real-robot data to enable more data-efficient online learning.

Paper

https://roboticsproceedings.org/rss18/p031.html

Code

https://github.com/UM-ARM-Lab/data-augmentation-for-manipulation

Video

https://www.youtube.com/watch?v=8j07Gy-1Zmc