Bias tree expansion using reinforcement learning for efficient motion planning모션계획을 위한 강화학습 기반 트리 편향 확장 기술

Abstract: Motion Planning is a computational problem to find a valid and optimal path from the given start to the goal configuration.
During the last few decades, sampling-based motion planning methods, such as Rapidly-exploring Random Tree* (RRT*), have been shown to work well even in a high-dimensional and continuous state space.
Recently, with the advances of deep learning, sample efficiency of sampling-based motion planning has been improved by learning the bias (heuristic) to the near-optimal region considering the surrounding obstacles and goal position.
However, the performance of a neural network trained using supervised learning is highly dependent on a set of demos previously collected for training.
Therefore, this leads to problems such as distribution mismatch and performance bounding and overfitting to the demos.
In this regard, we propose RL-RRT* to train the network using reinforcement learning and use it as a bias network.
We validate our method in a 2-D environment showing improved anytime performance, including initial solution quality and time and reasonably fast cost convergence rate.