This toolkit provides a framework for developing and comparing reinforcement learning algorithms within the ROS 2 and Gazebo ecosystem, targeting roboticists and ML researchers. It offers a real-world-application-oriented environment for benchmarking robot behaviors, building upon the concepts of OpenAI Gym.

How It Works

gym-gazebo2 integrates ROS 2 middleware with Gazebo simulations, enabling RL agents to interact with robotic systems. It leverages a standalone architecture, inspired by OpenAI Gym, to provide a structured interface for defining environments, actions, and observations. This approach facilitates seamless translation of learned policies between simulated and real MARA robotic arms.

Quick Start & Requirements

• Installation from source or via Docker (refer to INSTALL.md and docker/README.md).
• Requires ROS 2 and Gazebo.
• For Dashing, OpenSplice DDS implementation is recommended (export RMW_IMPLEMENTATION=rmw_opensplice_cpp).
• Example execution: cd ~/gym-gazebo2/examples/MARA && python3 gg_random.py -g.
• Documentation: https://acutronicrobotics.com/docs

Highlighted Details

• Supports both simulated and real MARA robotic arms.
• Offers command-line arguments for environment customization (e.g., --gzclient, --realSpeed).
• Allows running Gazebo without a visual interface for faster simulation speeds.
• Whitepaper available at https://arxiv.org/abs/1903.06278.

Maintenance & Community

• Initial release in March 2019.
• Further improvements to MARA environment observations released in March 2019.

Licensing & Compatibility

• The README does not explicitly state the license. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The project's ROS 2 compatibility notes mention potential bugs with Fast-RTPS in the Dashing distribution, recommending OpenSplice. The last update mentioned is from March 2019, suggesting potential maintenance gaps.