ADAPTIVE REINFORCEMENT LEARNING FOR SIMULATED ROBOTIC ARM MANIPULATION

Abstract— This research explores the use of adaptive

reinforcement learning for robotic arm manipulation with a Deep

Deterministic Policy Gradient (DDPG) algorithm. A 7-DOF robotic

arm was controlled using a MATLAB-based system integrated with

CoppeliaSim for simulation, employing a state machine for task

planning and execution. The DDPG algorithm trained an agent using

an actor-critic architecture to manage the continuous action space.

Over 500 episodes, the agent adapted to varying object properties and

tasks, achieving 75.40% accuracy and a 72.00% success rate. The

learning curve was sigmoidal, with an average reward of 24.12 per

episode. Q-value analysis indicated a preference for Lower and Place

actions. The average steps per episode (60.52) suggest efficiency

improvements are needed. The study highlights the need for better

exploration-exploitation balance and advanced techniques like meta

learning to enhance adaptability. Future work should optimize the

reward function, improve exploration strategies, and investigate

sophisticated algorithms for real-world applications.