Hybrid Learning-based Control for Reliable Robotic Arm Manipulation

Keya Sinha, Mandar Joshi, Avani Baranwal, Agniva Banerjee, Arijit Sen

Keywords: Robotic Arm Manipulators, Reinforcement learning, Hindsight Experience Replay, Model Predictive Control

Abstract: This study presents a hybrid learning-based control framework for robotic arm manipulation that handles sparse rewards and kinematic constraints by integrating Deep Reinforcement Learning (DRL) with Hindsight Experience Replay (HER) and Model Predictive Control (MPC) separately. The method pairs HER’s sample-efficient exploration with MPC’s trajectory optimization and safety within standard RL algorithms. Simulation results using the PandaReach-v3 environment demonstrate that the hybrid approach achieves a higher success rate and maintains constraint satisfaction with improved sample efficiency for robotic arm manipulators.

Submission Number: 14