Recent  advances  in the development  of sensors allowed to obtain robots with torque-sensitive sensors in each joint. At the moment,  these sensors are used only to detect collision. This work shows the possibility of obtaining information  on the collision point and  it type. This  information  can subsequently  be used to select the robot’s behavior  strategy. The  contact point localization  is realized  using two approaches: the analytical  approach and  machine  learning. Analytical  approach is based on finding point on the robot length and  direction of applied external  force where an equivalent  torques will be the same as torques in a real robot. In the machine  learning approach various learning technics were tested. For the collision type identification  a classification  tree was proposed that  distinguish  soft and  hard  collision,  purposeful  and  accidental, single and continuous.  The  algorithm at the first stage detects presence of a collision, and if there is a collision localizes it and identify  its type. The  described algorithms were tested on an industrial  manipulator  Kuka  iiwa LBR  14 R820, ground truth information  about the experiments  was obtained  using a 3D lidar.