Standstill Auto-Tuning of ADRC for Electro-Hydraulic Articulated Steering

Abstract

Manual tuning of the steering controller for an autonomous articulated vehicle is time-consuming and scales poorly across vehicle variants. This thesis proposes an automated pipeline that identifies a small set of plant parameters and uses them to configure an error-based Active Disturbance Rejection Control (ADRC) law, without manual calibration. The steering plant is described by three blocks, an input deadband, a second- order linear time-invariant (LTI) core, and a mechanical output saturation. The pipeline runs three steps in sequence at vehicle standstill. First, the input deadband width is identified offline using a breakpoint grid search, and then statically compensated. Second, a step is applied to the plant and a nonlinear least-squares estimator fits the gain K and time constant τ of the LTI core to the recorded response. Third, an empirical tuning rule built on top of bandwidth parameterisation and half-gain tuning turns K and τ into the ADRC controller and Extended State Observer gains, after which the observer estimates and cancels unmodelled dynamics online during operation. The pipeline is verified on three test environments, a set of second-order plants with known parameters, an OpenModelica multi-body simulation model of the vehicle, and the real articulated vehicle as the final end-to-end check.

Within the simulation environment, nine vehicle configurations span differences in chassis mass, cylinder bore, ground friction, valve flow capacity, and linkage geometry. Across these configurations the auto-tuned controllers achieved low tracking error, maintained control smoothness under measurement noise, and absorbed the shift in plant gain produced by payload and ground-friction changes away from the standstill operating point at which the parameters were identified.

On the real vehicle the identification stages ran end to end and returned a deadband and gain comparable to the simulated family, but the closed loop did not track the reference sweep. The plant model omits a transport delay, assumed negligible, which leaves the auto-tuned bandwidth too high for the real plant, and a measured deadband hysteresis the model also omits adds to the gap. Lowering the bandwidth helps but does not close it, and accounting for these effects is left as future work.