Kinodynamic Motion Planning for Robotic Arms Based on Learned Motion Primitives from Demonstrations

Learning from Demonstration (LfD) is a powerful tool for users to encode information about a task for a robot to perform. LfD has been used with some success in specific types of tasks, however very few implementations consider dynamic features in demonstrations while exploring new environments. The goal of this paper is to propose a novel motion planning algorithm that can incorporate the dynamics of a demonstration and avoid obstacles using learned motion primitives. The method uses a combination of hidden semi-Markov models (HSMM) and neural network controllers to classify and encode motion primitives and their sequences. The encoded motion primitives and their transition probabilities are then used to design a discrete sample space to be utilized by a random tree search algorithm. To evaluate this method, a bar-tending task that includes important dynamic motions was recorded. The recorded demonstrations were used in this method to create the discrete sample space and generate a trajectory for the task in a new environment. The algorithm was run 100 times with a randomly selected set of obstacles and found a feasible trajectory with 91% success.