MARA is a cool collaborative robot arm that was built using H-ROS components, meaning it has ROS-2 in its actuators, sensors, and any of the other representative modules inside it. Hence, each module can provide industrial-grade features, for example, ROS 2 features such as deterministic communication and component life cycle management. Due to the nature of H-ROS integration, the robot arm can be connected with additional sensors and actuators, as well as an easy system upgrade. The robot arm also supports a good number of industrial grippers. MARA is a 6-DoF robot arm with a payload capability of 3 kg and a tool speed of 1 m/s. The robot arm weighs about 21 kg and has a reach of 0.65 meters.

There are four environments in this reinforcement learning implementation:

• MARA: This is the simplest environment and is where the robot's tool center tries to reach a given point in space. The environment is reset if a collision is detected but the collision is not modeled with the reward function. The orientation of the tool is omitted as well.
• MARA Orient: The environment considers the translation and rotation of the end-effector of the robot and resets if a collision is detected. These are also not modeled into the reward system.
• MARA Collision: This environment is like the MARA environment (only translation), but with collisions modeled in the reward function. The robot arm receives a punishment if the robot arm collides and gets reset to the initial pose.
• MARA Collision Orient: This environment is a combination of MARA Collision and Orient, where the robot arm's translation and orientation are considered and the collision is modeled in the reward function. This is one of the most complex environment implementations in this package.

Let's see how to use these packages in action.