A Fast Planning Algorithm for Humanoids with Supervised Learning and Subproblem Decomposition

In this paper, we introduce a fast trajectory planning algorithm for humanoid robots. Our method uses direct transcription to formulate the motion optimization problem with centroidal dynamics and a full kinematics model. To enhance computational efficiency, our approach comprises two key components. Firstly, we leverage the supervised learning method to train a neural network that predicts reference trajectories for the optimization problem. The reference trajectory under a specific task is used as the initial guess to warm start the problem. Additionally, we decompose the original problem into subproblems. The subproblems contain all the constraints of the original problem. We iteratively solve these subproblems alternately until a convergent solution is achieved. Our approach enables the generation of a feasible motion planning solution in a short time with a high success rate. Simulation experiments under various motion scenarios confirm the validity of the algorithm.