Supersizing self-supervision: Learning to grasp from 50K tries and 700 robot hours

Abstract

Abstract-Current learning-based robot grasping approaches exploit human-labeled datasets for training the models. However, there are two problems with such a methodology: (a) since each object can be grasped in multiple ways, manually labeling grasp locations is not a trivial task; (b) human labeling is biased by semantics. While there have been attempts to train robots using trial-and-error experiments, the amount of data used in such experiments remains substantially low and hence makes the learner prone to over-fitting. In this paper, we take the leap of increasing the available training data to 40 times more than prior work, leading to a dataset size of 50K data points collected over 700 hours of robot grasping attempts. This allows us to train a Convolutional Neural Network (CNN) for the task of predicting grasp locations without severe overfitting. In our formulation, we recast the regression problem to an 18-way binary classification over image patches. We also present a multi-stage learning approach where a CNN trained in one stage is used to collect hard negatives in subsequent stages. Our experiments clearly show the benefit of using large-scale datasets (and multi-stage training) for the task of grasping. We also compare to several baselines and show state-of-the-art performance on generalization to unseen objects for grasping.

Fast grasp tool design: From force to form closure

“…Different strategies for grasp generation exist, for example analytical methods [6], leveraging large datasets for machine learning techniques [7] or sampling based approaches [8]. In the present work, we focus on sampling based techniques, since we strive to achieve fast generation of good grasps with a limited amount of data.…”

Preparing for the Unknown: Learning a Universal Policy with Online System Identification

“…While the results are impressive, these methods usually require extensive amount of experimental data [17,25] or relatively restrictive settings [16]. It is unclear whether these method would work directly on more dynamic motor skills in the real-world, such as locomotion.…”

“…In addition, progress has been made in directly learning neural network control policies of manipulation tasks for real robots [16,25,17]. While the results are impressive, these methods usually require extensive amount of experimental data [17,25] or relatively restrictive settings [16].…”

From Vision to Grasping: Adapting Visual Networks

“…Levine et al [9] train CNNs to learn a mapping straight from raw images to torques at the robot's motors, however they only use a few hundred examples to train a deep network. Pinto et al [11] focus on this problem collecting around 40,000 random trial and error grasp attempts to adapt a CNN based on AlexNet. Given an image patch, the output of their CNN predicts the likelihood of a successful planar grasp at the center of the patch in 18 discreet angles (0 • , 10 • , ... 170 • ), making it 18-way binary classification problem.…”