Dynamic Reward Shaping for Robotic Manipulation: A Self-Adjusting Mechanism Based on Task Complexity in Deep Reinforcement Learning

This paper presents a novel dynamic reward shaping framework for robotic manipulation tasks using deep reinforcement learning. Unlike traditional reward shaping methods, which use static functions, this approach adapts the reward structure in real-time based on task complexity indicators such as success rates, trajectory lengths, and exploration diversity. By modeling task complexity with Beta distributions, the system shifts from exploratory to exploitative learning as the agent becomes more proficient. The proposed mechanism enhances learning efficiency, stability, and scalability in high-dimensional robotic environments.