Research On Trajectory Planning Method Of 6-DOF Parallel Platform Based On Deep Reinforcement Learning

Parallel mechanism is a new type of manufacturing mechanism that has flourished in the world in recent years.It has special advantages such as good rigidity,high precision,high speed and acceleration.It has a broader application prospect in the fields of ship,aerospace,automobile,micro-production and medical treatment.However,there are some problems in the forward kinematics numerical method of parallel platform,such as the selection of iterative initial value.The positive solution algorithm based on neural network has the problems of low calculation accuracy and complicated combination with other intelligent algorithms.Traditional trajectory planning has problems such as low intelligence,low adaptability and complex planning process due to the need to spend a lot of time and resources to redesign polynomial spline curves in the face of different types of trajectory planning tasks and parallel platforms.In this paper,the 6-DOF parallel platform RSS mechanism is taken as the research object.Firstly,the 6PFK-ECDL algorithm is proposed.The algorithm inputs the joint angle into the residual neural network suitable for nonlinear regression.The network outputs a more accurate end pose,and then the end pose is iteratively calculated by error compensation to obtain a higher-precision positive solution result,which solves the problem of the above kinematics positive solution.The New Numerical algorithm and 6PFK-ECDL algorithm are verified by examples.The simulation results show that the average position error of 6PFK-ECDL algorithm is 55 % lower than that of 6PFK-6PFK-NN algorithm.The average attitude error is reduced by 63.9 %.The minimum calculation time is reduced by37.7 %.Then,aiming at the problems existing in the trajectory planning of the parallel platform at the present stage,the 6PTP-ELDDPG algorithm is proposed.The algorithm integrates the forward kinematics model based on the 6PFK-ECDL algorithm into the Deep Deterministic Policy Gradient algorithm.After one round of interaction with the environment,multiple dynamic planning is started.In each dynamic planning,the experience of the unfinished task in the experience pool is extracted.According to the current strategy and the state output action in this experience,this action is input into the forward kinematics model of the sixdegree-of-freedom parallel platform,and the reward value and the next state are obtained and stored in the experience pool,and the network is updated.Repeat this process until one round of interaction in dynamic planning is completed.Finally,the 6PTP-ELDDPG algorithm and Cartesian space linear trajectory planning algorithm are simulated and tested in real environment.The experimental results show that the convergence speed of 6PTP-ELDDPG algorithm is 40 % higher than that of 6PTPDDPG algorithm,the average number of steps to reach the target pose is reduced by 79.3 %,the success rate is increased by 42 % and the training is more stable.The average planning time of the 6PTP-ELDDPG algorithm is 33.4 % less than that of the Cartesian space straight line,the joint speed is more gentle,the joint acceleration is continuous and gentle,and the average joint maximum jerk is reduced by 72.5 %.There is no joint acceleration mutation problem in the Cartesian space straight line trajectory planning algorithm,and the trajectory that smoothly reaches the target pose can be planned in a relatively short time.