Lazy Steering RRT*: An Optimal Constrained Kinodynamic Neural Network Based Planner with no In-Exploration Steering

Abstract

Kinodynamic-RRT* provides a sampling based asymptotically-optimal solution for motion planning of kinematically- and dynamically-constrained robots. For nonlinear systems, normally, the time- and energy-clamped steering function solutions needed within the RRT* use numerical iterative schemes such as shooting methods, which are computationally cumbersome. The number of calls to these solvers increases with the size of the tree. Hence, the time complexity of finding feasible steering functions prevents kinodynamic-RRT* for non-linear systems from being utilized in realtime or in situations where fast planning and re-planning are needed. Kinematic/dynamic constraints reduction to make the steering functions solvable in real time has been proposed in literature, however, these methods would affect the optimality of the solution. In this paper, we propose a lazy-steering kinodynmaic RRT* in which, machine learning techniques are used to (1) predict if a randomly-selected node is steerable to, and (2) if the steering is deemed feasible, what would be the estimated energy cost associated, when executing it. This provides a promising framework for implementing Kinodynamic-RRT* in which the use of numerical methods is delayed (hence the name lazy steering) until a potential collision free path has been found, and only then the numerical techniques are invoked. This results in a huge improvement in the run time with little trade off on optimality. Our proposed method was tested via simulation for motion planning of an under-actuated, non-holonomic, quadcopter with nonlinear dynamics in an environment cluttered with obstacles. The lazy-steering RRT* was faster than its counterpart (which was based on some recent works) by two orders of magnitude.