Grasp Pose Sampling for Precision Grasp Types with Multi-fingered Robotic Hands

Generation of promising hand and finger poses for multi-fingered robotic hands cannot be simplified as the 2-dimensional model for grippers. Current approaches rely on heuristics that reduce the search space while ignoring a large number of candidates. We present a generative model that samples 6DoF poses for several types of precision grasps. Similarly to previous works, we start with a geometric heuristic to gather data. However, with a large enough samples we are able to sample grasp poses that are by a large margin more successful than using the heuristics. The model consists of 3 cascaded generative models that are based on the conditional Variational Auto-Encoder framework, and takes as input the desired grasp type, the object label, and the object's size. It generates a grasp posture, meaning the configuration of the fingers of the robotic hand, and a 6DoF pose. Our cascaded model samples first the finger joint configuration, followed by the Cartesian position of the object and finally the rotation of the object, our sampler divides the 6DoF in simpler problems, which lead to more successful grasps. In our experiments we show that our model improves the percentage of successful grasps sampled compared to the heuristic and compare several variants of the model to support our design choices, showing the benefits of the cascaded sampling.