TARG: Tree of Action-reward Generation With Large Language Model for Cabinet Opening Using Manipulator

In robotics, reinforcement learning (RL) is often used to help robots learn complex tasks through interactions with their environment. A crucial aspect of RL is the design of reward functions; these functions guide the learning process by providing feedback on a robot's actions. However, crafting these reward functions manually is time-consuming and requires extensive human expertise. In this paper, we propose a tree of action-reward generation (TARG) model that automates reward generation for a given task without the need for human fine-tuning. By using a large language model (LLM), we create a systematic action plan sequence to generate a tree of action that guides RL training. Proposed method facilitates the automatic generation of a reward tree, which stabilizes the training process. To demonstrate the effectiveness of the proposed TARG framework, we conducted experiments involving a cabinet opening task within the IsaacSim simulation environment. The results demonstrated the potential of the proposed framework to significantly improve the adaptability and performance of robots in complex settings.