Research

Automation of hydraulic excavators, with a focus on identification, estimation and control in free space and in contact with the ground. The work is part of the ENGAGE Doctoral Network (Marie Skłodowska-Curie Actions), is carried out in close collaboration with Novatron, and is developed and tested in EMC (Earthmoving Modelling and Control), our software for the simulation, identification and control of these machines.

Motion in free space

Adaptive joint-space and Cartesian control that learns the uncertain parts of the machine model while it moves, so the bucket follows its path more closely with every pass.

Four passes of a levelling cycle under an adaptive controller, with the tracking error ‖e‖ shrinking and the estimates θ̂ converging.

Into the ground

Impedance, admittance and force control change the planned path automatically when the bucket meets the ground, giving way where the soil pushes back and returning to the plan where it does not.

Admittance control along the path normal. Where the contact force exceeds F_d the executed path (blue) gives way by Δ over the hard layer, then returns to the plan.

Feeling what is buried

An estimate of the force at the bucket, from cylinder pressures and joint motion, reveals contact with buried pipes, cables or boulders so the dig can stop in time.

The force estimate F̂ leaves the band of the expected soil force at contact with a buried pipe; the dig stops and backs off.

Rocks and mixed material

Capturing rocks and digging mixed material, with the weight of a carried rock estimated from the hydraulic pressures alone.

Recursive estimation of a carried rock's mass m̂ with its ±2σ band; motion makes the load observable, standstill friction hides it.

The machine itself

The models behind the controllers, from cylinder friction and its static bounds to the mass and centre of mass of the machine, estimated from its own motion.

Cylinder friction against velocity with identified data, the fitted Stribeck curve and bounds on the static friction F_S.

Mass and centre of mass estimated while the boom moves, with the uncertainty shrinking and the stability margin d known.

Simulation

A simulator of the hydraulics and the mechanics together, from pump and valves to cylinders and linkage, where the controllers are designed and tested before they meet the machine.

One model from pump to load, with valve command u, chamber pressures p_A and p_B, and cylinder stroke x.

Earlier work

Earlier work covers robot learning during the M.Sc. and constrained control and fault diagnosis during the B.Sc. The full list of publications is in the CV.

Learning from demonstration

Robot programming by demonstration: segmentation and via-set optimization

I-RIM Conference 2025, pp. 205–207, 2025.

Several kinesthetic demonstrations are aligned and segmented; a small set of via-points is optimised and the arm reproduces the skill.

Demonstrations with different timing are aligned (event-based DTW), segmented by dynamic programming and summarised by a small set of via-points.

Constrained tracking control

Nonlinear tracking control algorithm for dynamical output systems manipulated by the hardly constrained oscillatory actuator

Structural Control and Health Monitoring, 2023, 3799011, 2023. Paper

A flexible arm tracks a reference through an actuator with hard limits; the input touches its bounds but never exceeds them.

Lyapunov design under input limits

Synthetic Lyapunov stabilization technique for designing actuation-constrained multi-input multi-output control systems

Journal of the Franklin Institute 359 (12), 5891–5918, 2022. Paper

A three-joint arm stabilised at a target with bounded torques; the Lyapunov function only decreases.

Rotor fault diagnosis

Pattern recognition of unbalanced rigid rotor bearing forces

Amirkabir Journal of Mechanical Engineering 54 (6), 1249–1270, 2022. Paper

Bearing forces of an unbalanced rigid rotor, computed from its model; once the unbalance is located (illustrated here, not a network run), a balancing mass removes it.