[Paper Review] SoftGym: Benchmarking Deep Reinforcement Learning for Deformable Object Manipulation

Xingyu Lin, Yufei Wang+2 more|arXiv (Cornell University)|Nov 14, 2020

Reinforcement Learning in Robotics53 references68 citations

TL;DR

This paper presents SoftGym, a benchmark for deformable object manipulation using deep RL, detailing environments, observation/action spaces, task variations, training/evaluation protocols, and baseline algorithm results. It analyzes image and feature observations across several tasks (water, cloth, rope) and compares methods like SAC, TD3, PlaNet, and RIG under different settings.

ABSTRACT

Manipulating deformable objects has long been a challenge in robotics due to its high dimensional state representation and complex dynamics. Recent success in deep reinforcement learning provides a promising direction for learning to manipulate deformable objects with data driven methods. However, existing reinforcement learning benchmarks only cover tasks with direct state observability and simple low-dimensional dynamics or with relatively simple image-based environments, such as those with rigid objects. In this paper, we present SoftGym, a set of open-source simulated benchmarks for manipulating deformable objects, with a standard OpenAI Gym API and a Python interface for creating new environments. Our benchmark will enable reproducible research in this important area. Further, we evaluate a variety of algorithms on these tasks and highlight challenges for reinforcement learning algorithms, including dealing with a state representation that has a high intrinsic dimensionality and is partially observable. The experiments and analysis indicate the strengths and limitations of existing methods in the context of deformable object manipulation that can help point the way forward for future methods development. Code and videos of the learned policies can be found on our project website.

Motivation & Objective

• Provide a modular, varied benchmark for deformable object manipulation to evaluate deep RL methods.
• Describe observation and action spaces for image-based and feature-based settings across multiple tasks.
• Introduce task variations to capture object size, shape, and physical property variations.
• Establish training/evaluation pipelines and baseline hyper-parameters for fair comparisons.

Proposed method

• Define two observation modalities (image 128x128x3 and task-specific feature vectors).
• Normalize actions to [-1,1] with environment-specific unnormalized ranges per task.
• Utilize a picker-based controller for cloth/rope tasks and a 1D cup motion for PassWater/PourWater.
• Pre-compute 1000 task variations per environment; use 800 for training and 200 for evaluation.
• Provide reference network architectures for image (CNN encoder, VAE/RIG specifics) and feature observations (MLPs).
• Present detailed hyper-parameters for SAC, TD3, PlaNet, and RIG across image/feature settings.

Experimental results

Research questions

• RQ1 How do image-based versus feature-based observations affect DRL performance on deformable object tasks?
• RQ2 What is the impact of task variations (size, shape, material properties) on learning and generalization?
• RQ3 How do different DRL algorithms (SAC, TD3, PlaNet, RIG) compare on SoftGym tasks under consistent training/evaluation protocols?
• RQ4 What observation and network architectures best support learning in deformable object manipulation?
• RQ5 What are the effects of full-state versus partial observations on learning efficiency and performance?

Key findings

• Image-based and feature-based observations yield different learning characteristics across tasks, with detailed architectures provided for each setting.
• TD3 with RGB observations showed poor learning in many experiments due to action saturation and Q-function bias issues.
• RIG and PlaNet employ latent representations and planning/HER-style techniques to tackle planning in high-dimensional observations.
• Full-state observations performed poorly for SAC/TD3 due to high dimension and redundancy compared to feature/image inputs.
• A large suite of task variations (cup sizes, rope twists, cloth sizes) was pre-computed to enable robust training and evaluation across environments.