Towards Autonomous Reinforcement Learning for Real-World Robotic Manipulation with Large Language Models