Energy-Efficient High-Dimensional Motion Planning for Humanoids Using Stochastic Optimization

This paper presents a method for planning a motion for a humanoid robot performing manipulation tasks in a high dimensional space such that the energy consumption of the robot is minimized. While sampling-based path planning algorithms, such as rapidly-exploring random tree (RRT) and its variants, have been highly effective for complex path planning problems, it is still difficult to find the minimum cost path in a high dimensional space since RRT-based algorithms extend a search tree locally, requiring a large number of samples to find a good solution. This paper presents an efficient nonmyopic motion planning algorithm for finding a minimum cost path by combining RRT* and cross entropy (CE). The proposed method constructs two RRT trees: the first tree is a standard RRT tree which is used to determine the nearest node in the tree to a randomly chosen point and the second tree contains the first tree with additional long extensions. By maintaining two separate trees, we can grow the search tree non-myopically to improve efficiency while ensuring the asymptotic optimality of RRT*. We first identify and demonstrate the limitation of RRT* when it is applied to energy-efficient path planning in a high dimensional space. Results from experiments show that the proposed method consistently achieves the lowest energy consumption against other algorithms.