Reinforcement Learning based End-to-End Control of Bimanual Robotic Coordination

To optimize the motion planning of multi robotic arms for a single task execution, traditional methods rely on manually planning the end effector's trajectory, which is time-consuming and labor-intensive. Reinforcement learning offers an alternative, where the robotic arm explores actions within its environment, receiving feedback through rewards, and learns control strategies for task completion. This research encompasses three primary areas: practical operation, simulation environment creation, and reinforcement learning algorithm training. For practical operations, the UR3 robotic arm model is programmed using a teaching device, facilitating motion planning via manual instruction. In simulation, a MuJoCo-based environment with a physics engine is developed to mirror the robotic arm's real-world movements. This simulation controls the arm through a mocap data stream, guiding the end effector's position and posture. Reinforcement learning training involves crafting various reward functions within the environment to achieve typical tasks, such as reaching targets and manipulating objects. Some preliminary results reveal that such a framework would facilitate the self-learning of multi agents in achieving the global target.