Target Localization for Robotic Arm Grasping Based on Deep Reinforcement Learning

Abstract

To address unstable target localization, insufficient end-effector alignment accuracy, and policy convergence difficulties when robotic arms perform grasping tasks in multi-distractor scenarios, this paper proposes a target localization method for robotic arms based on Proximal Policy Optimization (PPO). First, a PyBullet simulation platform is established, and a FAIRINO FR3 robotic arm simulation environment with random distractors is constructed. The target approach process of the robotic arm is modeled as a Markov Decision Process (MDP). Then, to reduce misalignment caused by distractors during target approach, a 10-dimensional state space with end-effector geometric center compensation is designed, and a composite reward function is optimized, including differential distance guidance, pose normal vector constraints, and a distractor false-hovering penalty. Finally, combined with the PPO algorithm using large-batch parallel sampling, the convergence and stability of the policy network are improved. Simulation results show that the proposed policy enables the robotic arm to align with the target object smoothly and stably while effectively reducing misalignment with random distractors, verifying its feasibility in random distractor scenarios.