Grasp Evaluation With Graspable Feature Matching

We present an algorithm that attempts to identify object locations suitable for grasping using a parallel gripper, locations which we refer to as “graspable features”. As sensor input, we use point clouds from a depth camera, from which we extract a gripper-sized voxel grid containing the potential grasp area and encoding occupied, empty, and unknown regions of space. The result is then matched against a large set of similar grids obtained from both graspable and non-graspable features computed and labeled using a simulator. While matching, we allow unknown regions of space in the input grid to match either empty or occupied space in our database of graspable feature voxel grids, which enables us to consider many possible geometries for occluded areas in our grasp evaluation. We evaluate our algorithm in simulation, using real-life sensor data of objects as input and evaluating the output using ground-truth object shape and pose.