Learning and Reacting with Inaccurate Prediction: Applications to Autonomous Excavation

Motivated by autonomous excavation, this work investigates solutions to a class of problem where disturbance prediction is critical to overcoming poor performance of a feedback controller, but where the disturbance prediction is intrinsically inaccurate. Poor feedback controller performance is related to a fundamental control problem: due to causality there is only a limited amount of disturbance rejection that feedback compensation can provide. It is known, however, that the use of predictive compensation can improve the disturbance rejection of a control system beyond the limitations of feedback. While prediction is desirable in a controller, the problem in excavation is that disturbance predictions are prone to error due to the variability and complexity of soil-tool interaction forces. This work proposes the use of iterative learning control to map the repetitive components of excavation forces into feedforward predictive compensation commands. Although experimental results show that preemptive disturbance compensation improves excavation performance, the non-repetitive nature of soil-tool interaction forces when digging is a source of inaccurate predictions. To explicitly address the use of imperfect predictive compensation, a disturbance observer is used to estimate the prediction error rather than the disturbance. To quantify inaccuracy in prediction, a feedforward model of excavation disturbances is interpreted as a communication channel that transmits corrupted disturbance previews, for which metrics based on the sensitivity function exist. During field trials the proposed method demonstrated the ability to iteratively achieve a desired dig geometry, independent of the initial feasibility of the excavation passes in relation to hydraulic actuator flow saturation. Under this iterative method predictive commands adapted to different soil conditions and passes were repeated autonomously until a pre-specified finish quality of the trench was achieved. Evidence of improvement in disturbance rejection is presented as a comparison of sensitivity functions of systems with and without the use of disturbance compensation given by feedforward action, also referred to as disturbance previews.