Comparison of Kinematic and Kinetic Reward Functions for RL for Robot Grasping

Robotic grasping is a complex challenge due to the interplay of perception, planning, and control, but learning-based approaches excel at handling high-dimensional gripper configurations, optimizing contact points, and adapting to dynamic environments. This work explores reinforcement learning (RL)-based robotic grasping using an anthropomorphic LEAP hand model. We investigate the effectiveness of different reward functions—kinematic (positional) and kinetic (contact force) feedback—on grasping performance. The work is implemented in the MuJoCo simulation environment, and policies are trained using the Proximal Policy Optimization (PPO) algorithm. While kinetic rewards demonstrate potential, the findings highlight limitations in current RL methodologies, such as susceptibility to overfitting and an inability to fully generalize. Future directions include integrating human feedback into reward functions and exploring model-based RL to address environment dynamics more comprehensively.