A method and system for detecting a collision by an autonomous robot based on a multi-sensor long short-term memory (LSTM) are disclosed. The method includes generating an input of an LSTM model by combining an input image received from an autonomous robot, light detection and ranging (LiDAR) distance information, and acceleration information, learning a collision alert situation by inputting the input to the LSTM model, and determining a collision situation using an output of the LSTM model and a fully connected neural network (FNN) model.