Closed-loop control dynamic obstacle avoidance algorithm based on a machine learning objective function

In this work, we address the issue of insufficient accuracy in the gradient projection algorithm and propose a closed-loop control dynamic obstacle avoidance algorithm that relies on a machine learning objective function. We initially establish a reasonable objective function and employ the gradient descent algorithm to enable dynamic obstacle avoidance in each bar. We then separate the end trajectory of the manipulator into multiple trajectory points and use the actual and expected positions of the end as the starting and ending points to significantly enhance the end tracking accuracy of the manipulator. Finally, we conduct simulation and real experiments on planar four degrees-of-freedom redundant manipulators to validate the efficacy of the algorithm. Moreover, the algorithm is proven to be applicable to dynamic obstacle avoidance under various trajectory tracking scenarios. It also exhibits advantages, such as smooth and continuous avoidance states and low computational costs.