Planning of Curvature-Optimal Smooth Paths for Industrial Robots Using Neural Networks

The use of industrial robots plays an increasingly important role in today’s production technology. Many fabrication processes like milling or gluing have high demands on accuracy and path smoothness to ensure good path tracking. Yet, industrial robot paths in cartesian space mainly consist of a sequence of linear and circular movements and hence show velocity and acceleration discontinuities in the segment transitions. This significantly limits the productivity and machining quality of the robot system. Traversing these discontinuities activates the kinematics mechanically and hence influences the accuracy and surface quality. In addition, the drives of the robot system are subjected to a higher load as a result which decreases the lifetime of the robot. Higher-order continuous paths are required to prevent this. Corner smoothing methods insert smooth, curvature continuous curves in the segment transition resulting in smooth paths. Methods based on polynomial smoothing splines either do not have any curvature-optimality properties or cannot be solved without violating the online constraints of the robot controller. To solve the conflict between the calculation of curvature-optimal smoothing curves and online execution in industrial robot controllers, this paper evaluates the use of neural networks as a model for calculating optimal geometry parameters for corner smoothing with polynomial splines. The model is trained applying supervised learning on a dataset containing offline generated pairs of geometries and optimal parameters. The presented method leads to curvature-optimized smoothing curves and is suitable for online path planning. The application of neural networks to this path planning problem is evaluated in a simulation using a digital twin model of an industrial robot.