Reactive Execution of Learned Tasks With Real-Time Collision Avoidance in a Dynamic Environment

This paper addresses the problem of learning from demonstration (LfD) and subsequent robot safety control in an unstructured dynamic environment different from the demonstrations. Generally, LfD has been successfully exploited for task programming, but the existing methods have not solved the problem of allowing the entire arm to avoid obstacles while satisfying the task motion constraints (e.g., the robotic arm approaching the target object while avoiding obstacles moving within the environment). To achieve this, we present an incremental LfD approach that combines a task-parameterized probabilistic model and the robot security domain to control a robot's behavior during task execution. Specifically, we propose a safety-oriented and task-oriented control strategy for redundant manipulators that makes full use of the motion redundancy of the manipulator and the space with no task restraints to satisfy the task constraints for human-robot coexistence. We then demonstrate the effectiveness of the proposed approach through a series of pick-and-pour experiments performed by a manipulator with 7 degree of freedom in a dynamic environment, where the robot must both avoid obstacles and satisfactorily complete the learned task with constraints.