Active learning of time-optimal motion for an industrial manipulator with variable payload

Generating optimal motion for robotic systems remains an active field of research, despite its relatively long history. In addition to being optimal with respect to a defined cost, such a motion must respect the electro-mechanical limitations of the system to be of use in real-world applications. Furthermore, in an industrial application where offline planning is not possible due to a priori unknown targets and payloads, the available time budget for computations is often highly limited, making many optimal motion planning approaches unsuitable. This work introduces a supervised learning-based motion planner for industrial manipulators that closely retains the superior performance of an optimal motion planner while requiring a fraction of the computations at runtime. The proposed motion planner implicitly considers the dynamics of the robot and therefore satisfies not only kinematic limits but also those of the actuators. The proposed method is compared to state-of-the-art approaches and demonstrated on a real robot.