Sim2Real Tools#
ManiSkill provides some useful tooling for doing sim2real training and deployment. This document discusses the Sim2RealEnv class and how it can help you perform sim2real experiments with less bugs and better sim2real alignment. It is intended for users who wish to do more advanced sim2real and/or build their own simulation environments to prepare for sim2real work.
We also have a simple tutorial that shows how to use visual Reinforcement Learning to train a cube picking policy in simulation using the low-costSO100 robot arm and 🤗 LeRobot system. At the end of the tutorial you will be able to deploy zero-shot a RGB policy successfully without needing any real world data (although it can help!). As there a lot of code that is often very specific to certain hardware and sim2real approaches, the tutorial an code for this particular sim2real approach is at our LeRobot-Sim2Real repository. The LeRobot-Sim2Real tutorial does not require you to follow any of the other tutorials and is self-contained.
Linked below are tutorials/docmentation on some tools and how to build with them. The Sim2RealEnv details how to align your simulated environment with the real world better using just a single wrapper. The Building Simulation Environments for Sim2Real outlines steps and details to follow to ensure the simulation environment is ready for sim2real (currently WIP).