SoftGym: Benchmarking Deep Reinforcement Learning for Deformable Object Manipulation

A deformable-object manipulation benchmark built on NVIDIA FleX, covering cloth, rope, and liquid tasks.

SoftGym is an open-source benchmark from Xingyu Lin and colleagues in David Held's group at Carnegie Mellon University, published at CoRL 2020, built specifically to evaluate reinforcement learning on deformable-object manipulation. Earlier RL benchmarks were mostly rigid-body or low-dimensional state tasks, whereas cloth, rope, and liquids have extremely high-dimensional state that is only partially observable — a completely different level of difficulty. SoftGym is built on NVIDIA's FleX particle physics engine (accessed through the Python interface PyFleX) and provides a standard OpenAI Gym interface. Tasks come in two tiers: medium difficulty includes transporting water, pouring water, straightening a rope, spreading cloth flat, and folding cloth; hard includes precise water pouring, folding a wrinkled cloth, dropping and then folding cloth, and shaping a rope into a specified form. The paper's experiments show existing algorithms still struggle noticeably on these tasks, especially when only given image observations. Because it depends on older CUDA and system versions, the authors recommend installing it via Docker.

ExampleIn SoftGym's FoldCloth task, an agent sees only an overhead camera image and controls two grasp points to fold a flat piece of cloth in half, rewarded by how well the two halves align.

Related

Deformable Object Manipulation · Garment Manipulation · Cloth Simulation · Fluid Simulation · Position-Based Dynamics · Benchmark

Sources

arXiv 2011.07215 - SoftGym
GitHub - Xingyu-Lin/softgym