This dataset contains the joint toque measurements of a robot manipulator (KUKA LWR4+) under accidental collisions and intentional contacts. It is specifically intended for the research study on robot collision detection, classification, diagnosis, or prediction. The dataset was recorded at Chair of Automatic Control Engineering, Technical University of Munich, Munich, Germany, by Dr. Zengjie Zhang, under the supervision of Dr. Dirk Wollherr, in 2017. Its detailed recording procedure is explained in the following work: [1] Zhang Z, Qian K, Schuller B W, and Wollherr D. An online robot collision detection and identification scheme by supervised learning and bayesian decision theory[J]. IEEE Transactions on Automation Science and Engineering, 2020, 18(3): 1144-1156. The dataset contains a number of external signal pieces of three classes: accidental collision (cls), with intentional manual contacts (ctc), and free from contacts (fre). Each signal piece lasts for 1.024s subject to the sampling rate 1kHz. Collisions or contacts occur at 0.256s of the signal pieces. The unit of the signal measurement is Nm. All the signals are recorded for the seven joints (#1 to #7) of the KUKA robot arm. The dataset is stored in .csv files. Each .csv file, containing the torque signal pieces for each class and each joint, is formed as an N by M matrix, where M = 1024 is the length of the signals and N is the number of signal pieces of the corresponding classes. For 'cls', N = 6960; for 'ctc', N = 7583; and for 'fre', N = 14098. Refer to the 'ReadMe.md' file for how to import the data to Python or MATLAB. This dataset is openly accessible for research work. Please cite this dataset and reference [1] if you publish the work based on them.