README.md

BioTac_slide_20_50

This repository contains tactile data and material description for texture classification, supplementary to the paper "Supervised Autoencoder Joint Learning on Heterogeneous TactileSensory Data: Improving Material Classification Performance", R. Gao et al.

We have included both the raw data and the compiled data here. The raw data includes a comprehensive range of modalities, provided by both tactile sensors and the robot arm. The compiled data includes purely the sensor data that are used in the abovementioned paper and have already been splitted in to train/validation/test sets.

Dataset description

This dataset concerns with 20 classes named with mat + x, where x is an index number. A snapshot of each corresponding material is shown below and the detailed description for each material can be found in material_properties text file by the name index or two-letter code.

Raw data set

Each material constitutes a separate folder in which 50 samples are presented, named by matx_n, where x is the material index and n is the sample index. For each sample, we have a series of .csv files compiled from original rosbag data and one .pt file that was used as input data for BioTac model mentioned in the above paper.

• For .csv files, the postfix after sample name indicates the recorded ros topic: bio-- BioTac sensor reading, cp-- cartesian position, cv-- cartesian velocity, cw-- cartesian wrench, jp-- joint position, jt-- joint torque, jv-- joint velocity. They are mostly used for checking the effectiveness of force control and velocity control for consistent data collection and can be further explored for future experiments.
• For .pt file, it is extracted from _bio.csv file for each sample and formatted as a 400*44 tensor, where 400 is the sequence length and 44 is the number of the data fields, as shown in the header of _bio.csv file. It can be directly imported by torch.load(filename) and used as input data for model training.

Compiled data set

The processed data set is presented to facilitate quick implementation in PyTorch. The data can be loaded with following commands

import torch
from tas_utils_bs import get_trainValLoader, get_testLoader
# data_dir = 'compiled_data/'
# kfold_number = 0 # number of fold that is chosen as the validation fold, range 0-3
# spike_ready = False # the data can be used for SNN training if the option is True
# batch_size = 32
# shuffle = True # where the data is shuffled for each epoch training 
train_loader, val_loader, train_dataset, val_dataset = get_trainValLoader(data_dir, k=kfold_number, spike_ready=False, batch_size=batch_size, shuffle=shuffle)
test_loader, test_dataset = get_testLoader(data_dir, spike_ready=False, batch_size=batch_size, shuffle=shuffle)

'''epoch and model setting here'''

for i, (XI, XB,  y) in enumerate(train_loader):
    print(XI.shape, XB.shape, y.shape) # torch.Size([32, 6, 10, 75]) torch.Size([32, 19, 400]) torch.Size([32])
    '''model training here'''