Barrier Offset Varying-Parameter Dynamic Learning Network for Solving Dual-Arms Human-Like Behavior Generation

Enabling robots to imitate human actions and perform tasks with high precision while avoiding potential obstacles in the environment can effectively enhance the interaction between social robots and humans. In this article, to achieve higher precision trajectory tracking and obstacle avoidance for dual-arm humanoid robots, the barrier offset varying-parameter dynamic learning neural (BOVDL) network method is proposed and applied to dual-arm humanoid behavior generation scheme. To do so, a dual-arms humanoid robot model is set up, and transformed into a constrained time-varying quadratic programming (TVQP) problem. Second, by using Lagrangian multiplier method and Karush–Kuhn–Tuchker condition, the inequality constrained TVQP is converted as a time-varying equation with a barrier parameter. Third, a varying-parameter dynamic learning network is presented to solve the time-varying equation with a barrier parameter. Computer simulation experiments are conducted to verify the feasibility, accuracy, and safety of the proposed BOVDL network method. Experimental results show that all 14 joints of the humanoid robot's arms are within the motion range of each real human arm's physical constraints. The maximum position error and velocity error between the desired trajectory and the actual trajectory of the end effector are less than $10^{-6}\ \text{m}$ magnitude and $10^{-7}\ \text{m}$ magnitude, respectively, representing a reduction of five orders of magnitude compared to the traditional varying-parameter convergent-differential neural network. Furthermore, the proposed method also enables the dual-arm humanoid robot to avoid collisions with obstacles while performing tasks, demonstrating the superiority of the proposed BOVDL network scheme.