©ISTOCKPHOTO.COM/ YUUJI By Qingkai Lu, Mark Van der Merwe, Balakumar Sundaralingam, and Tucker Hermans W e propose a novel approach to multifingered grasp planning that leverages learned deep neural network (DNN) models. We trained a voxel-based 3D convolutional neural network (CNN) to predict grasp-success probability as a function of both visual information of an object and grasp configuration. From this, we formulated grasp planning as inferring the grasp configuration that maximizes the probability of grasp success. In addition, we learned a prior over grasp configurations as a mixture-density network (MDN) conditioned on our voxel-based object representation. We show that this object-conditional prior improves grasp inference when used with the learned grasp success-prediction network compared to a learned, objectagnostic prior or an uninformed uniform prior. Our work is the first to directly plan high-quality multifingered grasps in configuration space using a DNN without the need of an external planner. We validated our inference method by Multifingered Grasp Planning via Inference in Deep Neural Networks Outperforming Sampling by Learning Differentiable Models Digital Object Identifier 10.1109/MRA.2020.2976322 Date of current version: 27 March 2020 1070-9932/20©2020IEEE JUNE 2020 * IEEE ROBOTICS & AUTOMATION MAGAZINE * 55http://www.ISTOCKPHOTO.COM/