MuJoCo_RL_UR5

Why we included this project

Robotic pick and place is usually trained on joint states, but this repository sets up the harder version: a MuJoCo Gym environment where the agent sees a camera image instead of a state vector, wrapping a UR5 arm and Robotiq gripper. That visual observation brings the setup closer to a real deployment, and the multi-discrete action space with a binary grasp-success reward is a concrete example of structuring a difficult manipulation task. The MJ_Controller class is useful on its own for basic arm control and for resetting the robot between episodes. This is a thesis companion project and no longer actively maintained, so treat the code as a teaching example or starting point rather than production software. For anyone prototyping grasp prediction or pick-and-place RL, it is a handy reference.