Deep Prediction of Swept Volume Geometries: Robots and Resolutions

Computation of the volume of space required for a robot to execute a sweeping motion from a start to a goal has long been identified as a critical primitive operation in both task and motion planning. However, swept volume computation is particularly challenging for multi-link robots with geometric complexity, e.g., manipulators, due to the non-linear geometry. While earlier work has shown that deep neural networks can approximate the swept volume quantity, a useful parameter in sampling-based planning, general network structures do not lend themselves to outputting geometries. In this paper we train and evaluate the learning of a deep neural network that predicts the swept volume geometry from pairs of robot configurations and outputs discretized voxel grids. We perform this training on a variety of robots from 6 to 16 degrees of freedom. We show that most errors in the prediction of the geometry lie within a distance of 3 voxels from the surface of the true geometry and it is possible to adjust the rates of different error types using a heuristic approach. We also show it is possible to train these networks at varying resolutions by training networks with up to 4x smaller grid resolution with errors remaining close to the boundary of the true swept volume geometry surface.