SoftGym: Benchmarking Deep Reinforcement Learning for Deformable Object Manipulation

  • 2020-11-14 03:46:59
  • Xingyu Lin, Yufei Wang, Jake Olkin, David Held
  • 5


Manipulating deformable objects has long been a challenge in robotics due toits high dimensional state representation and complex dynamics. Recent successin deep reinforcement learning provides a promising direction for learning tomanipulate deformable objects with data driven methods. However, existingreinforcement learning benchmarks only cover tasks with direct stateobservability and simple low-dimensional dynamics or with relatively simpleimage-based environments, such as those with rigid objects. In this paper, wepresent SoftGym, a set of open-source simulated benchmarks for manipulatingdeformable objects, with a standard OpenAI Gym API and a Python interface forcreating new environments. Our benchmark will enable reproducible research inthis important area. Further, we evaluate a variety of algorithms on thesetasks and highlight challenges for reinforcement learning algorithms, includingdealing with a state representation that has a high intrinsic dimensionalityand is partially observable. The experiments and analysis indicate thestrengths and limitations of existing methods in the context of deformableobject manipulation that can help point the way forward for future methodsdevelopment. Code and videos of the learned policies can be found on ourproject website.


Quick Read (beta)

loading the full paper ...