Adaptive residual filtering for safe human-robot collision detection under modeling uncertainties

This paper presents an innovative collision detection strategy for robot manipulators in the context of the human-robot interaction. Classical approaches consisting of a comparison of the applied motor torques with those provided by a dynamic model can be sensitive to model uncertainties, leading to conservative detection thresholds. In this work, a “gray-box” model is designed based on a use-case study to shape the on-line evaluation of the residuals. This approach takes into account unstructured uncertainties relative to the speed-dependent non-linearities (e.g. friction phenomena) and the acceleration, both of particular interest when dealing with highly time-varying dynamics. Taking advantage of proprioceptive measurements of the robot state, the residual is adaptively filtered regarding these model uncertainties, and the evaluation step is improved by considering a dynamic threshold. The proposed multi-variable algorithm is implemented on the CEA robot arm ASSIST and the experimental results illustrate the enhancement of the detection sensitivity.