Deep Reinforcement Learning for a Four Degree of Freedom Robot Arm Accelerated by Human Demonstration

This work investigates the use of deep reinforcement learning (DRL) for training a four-degree-of-freedom (4-DOF) robotic arm to efficiently and adaptively reach arbitrary target positions. A dueling double deep Q-network (D3QN) manages the large state-action space, while human demonstration data guides the agent's policy toward more human-like and effective trajectories. Experimental results show that integrating human demonstrations enables the robot arm to achieve more precise target positioning more frequently and in fewer steps than a purely self-trained agent. To assess generality and adaptability, the approach is tested on both a PhantomX Reactor robot arm and a Dobot Magician robot arm, with adjustments made to the Denavit-Hartenberg parameters of the latter. After running 5000 simulated training episodes on each platform, the resulting model is then implemented on the Dobot hardware for 50 episodes, demonstrating the feasibility of accurately controlling a robotic arm model different from the one used during training. The method's potential extends to practical computer vision-guided tasks such as pick-and-place or peg insertion, thus showcasing its versatility in both simulation and physical implementations.