Reinforcement Learning with Domain Randomization

Project Description

This repository includes the implementation of an algorithmic extension of previous adaptive domain randomization methods for the automatic inference of dynamics parameters for deformable objects. Based on the SofaGym platform, this work has been awarded with the Jury’s prize at the SOFA Week 2023.

GitHub repository

In this work, authors leverage the SOFA simulation platform to demonstrate how Domain Randomization (DR) enhances RL policies for soft robots. The approach improves robustness against unknown dynamics parameters and drastically reduces training time by using simplified dynamic models. They introduce an algorithmic extension for offline adaptive domain randomization (RF-DROPO) to facilitate sim-to-real transfer of soft-robot policies. Their method accurately infers complex dynamics parameters and trains robust policies that transfer to the target domain, especially for contact-reach tasks like cube manipulation.

All DR-compatible benchmark tasks and their method’s implementation are accessible as a user-friendly extension of the SofaGym framework. This software toolkit includes essential elements for applying Domain Randomization to any SOFA scene within a Gym environment, using the Stable Baselines3 (SB3) library for Reinforcement Learning training, allowing for the creation of multiparametric SOFA scenes and training control policies capable of achieving Sim2Real transfer. Example scenes are provided to guide users in effectively incorporating SOFA simulations and training learning algorithms.

Related publications

This repository contains the code for the paper “Domain Randomization for Robust, Affordable and Effective Closed-loop Control of Soft Robots – Gabriele Tiboni, Andrea Protopapa, Tatiana Tommasi, Giuseppe Averta – IROS2023.

Additional links

Read more on the dedicated website.

Back to the plugin list

Related Projects

Back to Top