Deep Koopman MPC Brings Real-Time Control to Wheel-Loader Cycles

As reported by arXiv (cs.RO), researchers have proposed a new control framework for automating the repetitive "V-cycle" maneuvers wheel loaders perform during earthmoving work—the forward dig-and-scoop, then reverse-and-dump sequence operators repeat throughout a shift.

The team, led by Armin Abdolmohammadi, points out that the nonlinear dynamics of articulated vehicles, combined with complex vehicle-terrain interactions, have limited how well conventional model-based controllers can handle these maneuvers. Their proposed solution is a hierarchical framework that combines long-horizon geometric planning with data-driven predictive control.

A reduced-order articulated kinematic model first generates the overall maneuver geometry, jointly optimizing the forward and reverse trajectories around a shared intermediate state. To capture the vehicle's actual dynamics, the researchers trained two deep bilinear Koopman models—one for forward motion, one for reverse—using data generated in Algoryx Dynamics, a high-fidelity physics simulator. Koopman-based methods approximate complex nonlinear dynamics using more tractable, near-linear representations, making them cheaper to evaluate inside real-time controllers.

These learned Koopman models feed into a model predictive control (MPC) system for trajectory tracking, engineered to run efficiently enough to execute within a 50-millisecond control loop—fast enough for real-time operation on heavy machinery.

According to the paper, high-fidelity simulation results show the end-to-end framework can execute wheel-loader V-cycle maneuvers accurately while remaining computationally efficient, offering what the authors describe as "a promising approach toward autonomous operation of articulated heavy-duty machinery."

The paper, submitted September 22, 2026, spans 8 pages with 4 figures and is posted to arXiv under the Robotics and Systems and Control categories. As with most arXiv preprints, the results are demonstrated only in simulation rather than on physical hardware, and the work does not appear to have completed formal peer review.