Collision Detection for Robotic Manipulators using LSTM Autoencoder

As robot collaboration scenarios become more prevalent, ensuring operational safety is increasingly critical. Collision detection for robotic manipulators is a key metric for ensuring system safety. This paper proposes a collision detection method based on an LSTM autoencoder anomaly detection approach. First, the estimated external torque is calculated with generalized momentum observer. Then, the estimated value is input into the network alongside sensor data. This approach relies solely on data collected under normal operating conditions, eliminating the need for collision scenario simulation.The experiment involving the six-degree-of-freedom robot manipulator achieved precision and recall rates of 92% and 97% respectively, indicating that this method can effectively identify collisions as abnormal events.