2015/09/22 by Lerrel Pinto, Abhinav Gupta, Pinto, Lerrel +1 · 14 citations
Computer Science · Engineering · #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG) #Reinforcement Learning in Robotics #Robot Manipulation and Learning #Robotics (cs.RO) #Soft Robotics and Applications
paper · pdf · doi:10.48550/arxiv.1509.06825
openalex publication_date 2015/09/22 · openalex created_date 2022/10/02 · openalex updated_date 2026/07/28
Current learning-based robot grasping approaches exploit human-labeled\ndatasets for training the models. However, there are two problems with such a\nmethodology: (a) since each object can be grasped in multiple ways, manually\nlabeling grasp locations is not a trivial task; (b) human labeling is biased by\nsemantics. While there have been attempts to train robots using trial-and-error\nexperiments, the amount of data used in such experiments remains substantially\nlow and hence makes the learner prone to over-fitting. In this paper, we take\nthe leap of increasing the available training data to 40 times more than prior\nwork, leading to a dataset size of 50K data points collected over 700 hours of\nrobot grasping attempts. This allows us to train a Convolutional Neural Network\n(CNN) for the task of predicting grasp locations without severe overfitting. In\nour formulation, we recast the regression problem to an 18-way binary\nclassification over image patches. We also present a multi-stage learning\napproach where a CNN trained in one stage is used to collect hard negatives in\nsubsequent stages. Our experiments clearly show the benefit of using\nlarge-scale datasets (and multi-stage training) for the task of grasping. We\nalso compare to several baselines and show state-of-the-art performance on\ngeneralization to unseen objects for grasping.\n