Determining Sample Quantity for Robot Vision-to-Motion Cloth Flattening

It is not clear how much data is needed to train a neural network efficiently. The “more is better” approach is not necessarily true for a multitude of reasons, especially when data is expensive to obtain in terms of time and effort, as is the case with robots in the real world. In this paper, we describe a method for determining the sample quantity for training a vision-to-motion DNN (Deep Neural Network). The training task is to straighten a single crumpled corner on a white towel in a simulated environment. For this purpose, the neural network was trained to identify a correct grab point for the simulated fabric as well as the direction and length of the movement to straighten the towel. We have trained it multiple times with datasets differing in the amount of samples. Comparisons show that the results hardly improve above a certain number of samples. In our experiments, datasets with 1000 and 2000 samples proved to be the most promising by striking a good balance between training time and desired accuracy. These models had a median error in predicting grab point position of 1.095 px and 0.977 px, respectively. Both reached or just exceeded the median error threshold of 1 px, which corresponds to an error of about 2.4 mm in the simulated environment. The study serves as a basis for transferring the movement to the real world, where Sim-2-Real approaches must be employed. Creating a dataset using only real world data with these sample quantities is not feasible.