SoftGym

SoftGym is a set of benchmark environments for deformable object manipulation including tasks involving fluid, cloth and rope. It is built on top of the Nvidia FleX simulator and has standard Gym API for interaction with RL agents.

Using Docker

The provided Dockerfile is based on the pre-built image for softgym which in turn uses CUDA 9.2. This codebase is tested with Ubuntu 20.04 LTS and Nvidia driver version 450.102.04

Prerequisites

• Install docker-ce
• Install nvidia-docker (Docker versions earlier than 19.03 require nvidia-docker2 and the --runtime=nvidia flag. On versions including and after 19.03, you will use the nvidia-container-toolkit package and the --gpus all flag)

Building and running Dockerfile

1. run make build to build the docker image
2. run make bash to run the image with an interactive bash session (step 2 also includes step 1)
3. run source setup.sh to setup the softgym conda environment and compile the python bindings for pyflex

SoftGym Environments

Image	Name	Description
	DropCloth	Lay a piece of cloth in the air flat on the floor
	FoldCloth	Fold a piece of flattened cloth in half
	SpreadCloth	Spread a crumpled cloth on the floor
	PourWater	Pour a cup of water into a target cup
	TransportWater	Move a cup of water to a target position as fast as possible without spilling out the water
	StraightenRope	Straighten a rope starting from a random configuration
	PourWaterAmount	This task is similar to PourWater but requires a specific amount of water poured into the target cup. The required water level is indicated by a red line.
	FoldCrumpledCloth	This task is similar to FoldCloth but the cloth is initially crumpled
	DropFoldCloth	This task has the same initial state as DropCloth but requires the agent to fold the cloth instead of just laying it on the ground
	RopeConfiguration	This task is similar to StraightenCloth but the agent needs to manipulate the rope into a specific configuration from different starting locations.

To have a quick view of different tasks listed in the paper (with random actions), run the following commands: For SoftGym-Medium:

• TransportWater: python examples/random_env.py --env_name PassWater
• PourWater: python examples/random_env.py --env_name PourWater
• StraightenRope: python examples/random_env.py --env_name RopeFlatten
• SpreadCloth: python examples/random_env.py --env_name ClothFlatten
• FoldCloth: python examples/random_env.py --env_name ClothFold
• DropCloth: python examples/random_env.py --env_name ClothDrop

For SoftGym-Hard:

• PourWaterAmount: python examples/random_env.py --env_name PourWaterAmount
• FoldCrumpledCloth: python examples/random_env.py --env_name ClothFoldCrumpled
• DropFoldCloth: python examples/random_env.py --env_name ClothFoldDrop
• RopeConfiguration: python examples/random_env.py --env_name RopeConfiguration

Please refer to softgym/registered_env.py for the default parameters and source code files for each of these environments.

Cite

If you find this codebase useful in your research, please consider citing:

@inproceedings{corl2020softgym,
 title={SoftGym: Benchmarking Deep Reinforcement Learning for Deformable Object Manipulation},
 author={Lin, Xingyu and Wang, Yufei and Olkin, Jake and Held, David},
 booktitle={Conference on Robot Learning},
 year={2020}
}

References

• NVIDIA FleX - 1.2.0: https://github.com/NVIDIAGameWorks/FleX
• Our python interface builds on top of PyFleX: https://github.com/YunzhuLi/PyFleX