Learning Probabilistic Generative Models For Fast Sampling-Based Planning

Due to their simplicity and efficiency in high dimensional space, sampling-based motion planners have been gaining interest for robotic manipulation in recent years. We present several new learning approaches using probabilistic generative models for fast sampling-based planning. First, we propose fast collision detection in high dimensional configuration spaces based on Gaussian Mixture Models (GMMs) for Rapidly-exploring Random Trees (RRT). In addition, we introduce a new probabilistically safe local steering primitive based on the probabilistic model. Our local steering procedure is based on a new notion of a convex probabilistically safety corridor that is constructed around a configuration using tangent hyperplanes of confidence ellipsoids of GMMs learned from prior collision history. For efficient sampling, we suggest a sampling method with a learned Q-function with linear function approximation based on feature representations such as Radial Basis Functions. This sampling method chooses the optimal node from which to extend the search tree via the softmax function of learned state values. We also discuss a novel constrained sampling-based motion planning method for grasp and transport tasks with redundant robotic manipulators, which allows the best grasp configuration and approach direction to be automatically determined. Since these approaches with the learned probabilistic models require large size data and time for training, it is essential that they are able to be adapted to environmental change in an online manner. The suggested online learning approach with the Dirichlet Process Mixture Model (DPMM) can adapt the complexity to the data and learn new Gaussian clusters with streaming data in newly explored areas without batch learning. We have applied these approaches in a number of robot arm planning scenarios and have shown their utility and effectiveness in simulation and on a physical 7-DoF robot manipulator