Optimization Algorithm for Robotic Hand Grasping Strategy Towards Multi-Morphology Targets

Robotic hand grasping of multi-morphology targets presents significant challenges in robotics due to diverse object shapes and varying grasping scenarios. Traditional reinforcement learning algorithms suffer from convergence instability and low sample efficiency. This study proposes a Smooth Clip Advantage Proximal Policy Optimization (SCAPPO) algorithm specifically designed for multi-morphology grasping tasks. SCAPPO addresses conventional PPO limitations by introducing a sigmoidbased smooth clipping function and an advantage prediction loss function to mitigate overfitting. Additionally, a segmented asymptotic reward function tackles the sparse reward problem in grasping tasks. Extensive experiments in PyBullet simulation environment using a multi-fingered robotic hand validate the algorithm's effectiveness. Results demonstrate significant improvements in grasping success rate, convergence speed, and adaptability compared to traditional PPO, DDPG, and TRPO methods. SCAPPO successfully handles various object morphologies including spherical, cylindrical, cubical, and irregular shapes, achieving an average grasping success rate of 89.7% across different target categories, providing an effective solution for complex multi-morphology grasping scenarios.