RLBench is a large-scale benchmark and learning environment for robot manipulation research, focusing on vision-guided tasks like reinforcement learning, imitation learning, and few-shot learning. It provides a flexible framework for researchers to develop and evaluate algorithms in a simulated robotics setting.

How It Works

RLBench leverages CoppeliaSim for physics simulation and PyRep as a Python API to interact with the simulator. It offers a wide variety of pre-defined manipulation tasks and supports custom task creation. The environment is designed to facilitate research in areas such as few-shot learning through curated task sets and sim-to-real transfer via domain randomization capabilities.

Quick Start & Requirements

• Install: pip install git+https://github.com/stepjam/RLBench.git
• Prerequisites: CoppeliaSim v4.1.0 (download and set COPPELIASIM_ROOT environment variable).
• Headless Rendering: Requires X server configuration (nvidia-xconfig, X :99 &).
• Documentation: Website and Paper

Highlighted Details

• Supports multiple robot arms (Franka Panda, Mico, Jaco, Sawyer, UR5) and grippers.
• Includes Gym compatibility for seamless integration with existing RL toolkits.
• Offers pre-defined task sets for few-shot and multi-task learning research.
• Provides tools and tutorials for creating custom tasks.

Maintenance & Community

• Active development with version 1.2.0 released in Feb 2022.
• Discord channel available for community support.
• Discord

Licensing & Compatibility

• License: Not explicitly stated in the README.
• Compatibility: Designed for research; commercial use implications are unclear without a specified license.

Limitations & Caveats

The README advises caution when using low-dimensional task observations, as they may not capture critical state changes (e.g., object slipping) that image-based observations would. It also notes that changing default image observation sizes may require re-collecting demonstrations.