Abstract: The unknown payload will increase the error of the robot model, which will cause serious interference to the collision detection and even lead to algorithm failure. Therefore, a collision detection algorithm for robots working with payloads is proposed in this paper. Firstly, a novel torque observer with band-pass filtering characteristics is designed based on the generalized momentum of the robot. Through the analysis of the frequency domain characteristics of the robot dynamics, the maximum frequency threshold of the robot dynamics model corresponding to the joint speed is obtained, and the parameters of the band-pass torque observer are determined. Then, the band-pass torque observer is used to filter the joint torque signal to improve the tolerance of the collision detection algorithm to the robot model error. Finally, the effectiveness of the proposed collision detection algorithm is further validated by human-robot interaction experiments. The experimental results show that the proposed algorithm is more robust than existing algorithms, and can meet the needs of the robot with 0～2 kg unknown payload.