RRT-Based Path Planning Considering Initial and Final Pose under Curvature Constraints for Nonholonomic Wheeled Robot

Abstract

Rapidly-exploring random trees(RRT) are popular algorithms in path planning, because they provide efficient solutions to singe-query problems and possess probabilistic completeness. Its modifications, such as RRT* and Informed RRT*, extend RRT to asymptotically find optimal solutions as the number of sampling approaches infinity. These algorithms, however, give no considerations to robot's poses and kinematic constraints, and therefore their results can be unfeasible for a nonholonomic robot with given initial pose and desired final pose. In this paper, we present another modification of RRT* for nonholonomic path-planning with not only kinematic constraints but also initial and final pose constraints. The proposed method constructs the search tree as a directed graph of which each node retains the position (x, y) plus the robot pose $\theta$ , and a segment of clothoid curve and line is used for connecting and evaluating the cost between two nodes. Furthermore, by building two trees from both initial and final poses and executing bidirectional search, this method can find a path containing crosscut point. We experimentally show that our approach calculates smoother, more feasible paths than RRT* and satisfy the given constraints on curvature.