Adaptive Filtering for Robust Proprioceptive Robot Impact Detection Under Model Uncertainties

Abstract

In the context of safe human-robot physical interaction, this paper introduces a new method for the detection of dynamic impacts of flexible-joint robot manipulators with their environment. The objective is to detect external impacts applied to the robot using only proprioceptive information with maximal sensitivity. Several model-based detection methods in robotics are based on the difference, called residual, between the estimated and the actual applied torques. Sensitivity of such methods can be limited by model uncertainties that originate either from errors on experimentally identified model parameters, possibly varying with the operating conditions, or the use of simplified models, which results in a residual dependence on the robot's state. The main contribution of this paper consists of a new adaptive residual evaluation method that takes into account this dependence, which otherwise can lead to a tradeoff between sensitivity and false alarm rate. The proposed approach uses only proprioceptive motor-side measurements and does not require any additional joint position sensors or force/torque sensors. Dynamic effects of a collision on the residual are isolated using bandpass filtering and comparison with a state-dependent dynamic threshold. Adaptive online estimation of filter coefficients avoids the need for extensive experiments for parametric model identification. Experimental evaluation on the CEA backdrivable ASSIST robot arm illustrates the enhancement of the detection sensitivity.