Closed-loop global motion planning for reactive execution of learned tasks

We present a motion planning approach for performing a learned task while avoiding obstacles and reacting to the movement of task-relevant objects. We employ a closed-loop sampling-based motion planner that acquires new sensor information, generates new collision-free plans that are based on a learned task model, and replans at an average rate of more than 10 times per second for a 7-DOF manipulator. The task model is learned from expert demonstrations prior to task execution and is represented as a hidden Markov model. During task execution, our motion planner quickly searches in the Cartesian product of the task model and a probabilistic roadmap for a plan with features most similar to the demonstrations given the locations of the task-relevant objects. We improve the replan rate by using a fast bidirectional search and by biasing the sampling distribution using information from the learned task model to construct high-quality roadmaps. We illustrate the efficacy of our approach by performing a simulated navigation task with a 2D point robot and a physical powder transfer task with the Baxter robot.