In this paper, we examines the integration of LLMs in designing reward functions for reinforcement learning (RL) to enhance robotic applications with minimal human input. In RL, the reward function is pivotal, guiding the agent’s learning trajectory by evaluating the desirability of behaviors within specific environments. Traditional reward functions, often sparse, lead to slow convergence as agents require extensive interactions to learn effectively. By leveraging LLM’s ability to generate code from task semantics, we propose a new method that reduces the complexity of reward design, allowing even non-experts to create effective reward policies using semantic prompts. We utilize the Soft Actor-Critic (SAC) algorithm, known for its efficiency and stability, to train agents under these conditions. To validate the efficacy of our method, we compare it with traditional techniques like Trajectory-ranked reward extrapolation (T-REX). Our findings indicate that the LLM-generated rewards enable quicker convergence and are as effective as those crafted through conventional methods, demonstrating the potential of LLMs to revolutionize reward shaping in RL. Furthermore, we transferred the robot door-opening task from the real-world simulation environment to a real robot, achieving sim-to-real.This approach allows for the rapid deployment of robotic systems, making sophisticated robotics technology more accessible and feasible for a wider range of applications. This study underscores the transformative impact of integrating advanced language models into the realm of robotics and RL, opening up new avenues for future research and application.