Learning-Based Real-Time Detection of Robot Collisions Without Joint Torque Sensors
Kyu Min Park, Jihwan Kim, Jinhyuk Park, Frank C. Park
Abstract
Robots operating in close proximity to humans require fast and reliable detection of collisions, which can range from sharp impacts (hard collisions) to pulling-pushing-catching motions (soft collisions). Because joint torque sensors can be costly, the external joint torques caused by collisions are typically estimated from motor current measurements at the joint actuators. For reliable detection, however, these methods require accurate models of dissipative friction at the joints. In this letter we design two learning-based detection methods - a support vector machine (SVM), and a convolutional neural network (CNN) - that require only motor current measurements together with a robot dynamics model and a momentum observer; no friction model is needed, and manual tuning of collision detection thresholds for each of the joints can be avoided. Extensive experiments with a six-dof industrial collaborative robot trained with stacks of four- and six-dimensional time series data validate our algorithms for a wide range of hard and soft collisions. The CNN-based method shows better performance if more training data is available, whereas the SVM-based method performs better with less data.