Physics-Guided Machine Learning Approach to Safe Quasi-Static Impact Situations in Human–Robot Collaboration

Following the performance and force limitation method of the ISO/TS 15066 standard, safety of a human–robot collaboration task is assessed for critical situations assuming quasi-static impact. To this end, impact forces and pressures are experimentally measured and compared with limit values specified by ISO/TS 15066. Consequently, such a safety assessment must be repeated whenever something changes in the collaborative workspace or the task, which severely limits the flexibility of collaborative systems. To overcome this problem, in this paper, a physics-guided machine learning (ML) method for prediction of peak impact forces, within predefined modification dimensions of collaborative applications, is proposed. Along with a pose-dependent linearized model, an ensemble of boosted decision tree (BDT) in combination with a feed-forward neural network (NN) is trained with peak impact forces measured at a UR10e robot covering the range of interest. A generic pick and place task with two modification dimensions is considered as an example of the presented methodology. The method yields the maximal safe impact velocity in the collaborative workspace.