Abdullah Ahmed Abdelmoez Mohamed

Grasping delicate fruits, such as berries or peaches, poses significant challenges in robotic manipulation due to their soft, deformable nature and susceptibility to damage. This propsal proposes the application of Diffusion Policy, a novel reinforcement learning framework based on diffusion models, to address these challenges. By modeling action sequences as a conditional generative process, Diffusion Policy makes use of multimodal sensory inputs, including vision and tactile feedback, to generate precise and adaptive grasping actions. The approach excels in handling the complex, high-dimensional action spaces required for delicate fruit grasping, which will be tested in both simulation and real-world environments. A comparison will be made between Diffusion Policy and traditional reinforcement learning methods, such as Soft Actor-Critic (SAC) and Deep Deterministic Policy Gradient (DDPG) in simulation and in cluttered real-world settings. Key advantages include Diffusion Policy ability to generalize to various fruit shapes and textures, as well as its robustness in dynamic environments. However, challenges such as computational complexity and sim-to-real transfer will be addressed in the future. This work highlights the potential of Diffusion Policy for delicate fruit grasping in agricultural and food processing applications, with future directions focusing on optimizing real-time performance and integrating advanced tactile sensors.