Learning-based Adaptive Sampling for Manipulator Motion Planning

Fast generation of optimized robot motions is crucial for achieving fluent cooperation in shared workspaces. Established sampling-based motion planning algorithms are guaranteed to converge to an optimal solution but often deliver low-quality initial results. To this end, learning-based methods reduce planning time delays and increase motion quality. Existing methods show promising results for low-dimensional and simulated problems. In the real world, sensor noise or a change of the robot’s tool can cause a distributional shift to the training data. An adaptive sampling strategy is thus required to cope with possibly suboptimal samples and ensure fast motion planning in human-robot collaboration. In this work, we present a sampling strategy for fast and efficient manipulator motion planning which is based on a conditional variational autoen-coder. We test our model for three optimization objectives: path length in configuration space and workspace, as well as joint limit distances. In contrast to other works, we not only condition our model on the planning problem but also on motion progress. This allows for generating samples in the growth direction of the tree. Using our method, we obtain high-quality initial paths within less than one second of planning time.