Collision Detection for Robot Manipulators Using Unsupervised Anomaly Detection Algorithms

Abstract

Collaborative robot manipulators operating in dynamic and unstructured environments shared with humans require fast and accurate detection of collisions. When using model-based detection methods without joint torque sensors, proper treatment for friction in the motors is required, such as accurate modeling of friction parameters. In our previous work, we have presented two supervised learning-based detection methods, which successfully compensate for uncertain dynamic effects, including unmodeled friction. However, robot collision data are required for training and only the scenarios that are learned as collisions can be robustly detected. In this article, we present two collision detection methods using unsupervised anomaly detection algorithms-a one-class support vector machine, and an autoencoder-for robot manipulators. Only the motor current measurements together with a robot dynamics model are required while no additional external sensors, friction modeling, or manual tuning of multiple detection thresholds are needed. Extensive experiments with a robot collision dataset collected with a six-joint robot manipulator validate that our unsupervised detection algorithms can robustly detect a variety of hard and soft collisions with very light computation using only the data for collision-free motions and without the knowledge of every possible type of collision that can occur.