Research on obstacle avoidance motion planning of robotic arm based on RG-DQN

In contemporary engineering applications, obstacle avoidance planning for robotic manipulators remains a critical challenge. The Deep Q-Network (DQN) algorithm is prone to becoming trapped in local minima easily during the training process. CMN-RRT(Central Multi-Node RRT) guided Deep Q-Network (RG-DQN) is proposed in this paper to address the common issue of local minima in obstacle avoidance problem of the robotic arm using traditional DQN algorithms. The integration of CMN-RRT enhances the agent’s training efficiency, resulting in reduced training time and faster decision-making during execution. In this study, both the RG-DQN and traditional DQN algorithms are used to train the KINOVA JACO GEN2 robotic arm in simulated working environment. The results show that in complex environments, the RG-DQN algorithm reduced total training steps by 23.3% and obtain 20.5% improvement in average reward values during the final 25 training episodes. Simulation results also demonstrate that RG-DQN enables higher quality obstacle avoidance training in a shorter time frame. In path planning tasks, the RG-DQN-trained agent outperformed the traditional CMN-RRT algorithm, reducing planning time by 39.2% and path length by 5.1%.