关键词: 轨迹优化, 物理约束, 自适应深度神经网络, 时间-冲击最优, 多目标优化

Abstract: Due to the unreasonable motion trajectory planning of the robotic arm,which fails to effectively balance the motion efficiency and stability,a trajectory planning method for the robotic arm based on the Physics-Informed Adaptive Deep Neural Network (PI-ADNN) was proposed. Firstly,a non-uniform B-spline curve was used to construct the motion trajectory model,and a multi-objective optimization model with the goals of time optimization and minimum impact was established. Secondly,a physical constraint adaptive neural network was designed,and the kinematic and dynamic constraints were introduced into the network loss function. Through the adaptive weighting mechanism,the balance between time and impact was achieved. Finally,simulations and experiments were conducted using the UR5 robotic arm,and comparisons were made with the commonly used Non-dominated Sorting Genetic Algorithm Ⅱ (NSGA-Ⅱ) and Improved Multi-Objective Evolutionary Algorithm (IMOEA). The simulation results show that PI-ADNN achieves the shortest motion time. Compared with the benchmark methods NSGA-Ⅱ and IMOEA, the motion efficiency of the method proposed is improved by 31.6 % and 18.8 %,respectively,and the jerk curves are smoother,with no jitter or impact occurring during the manipulator motion. The experimental results on the UR5 prototype further demonstrate that all joints of the robotic arm can maintain stable motion with a reduced motion time,thereby verifying the feasibility of PI-ADNN for manipulator trajectory optimization.

Key words: trajectory optimization, physical constraints, adaptive deep neural network, optimal time-jerk, multi-objective optimization