Researchers at the University of California, Davis programmed a simulated wheel loader to drive toward a material pile, reverse to a switchback point and approach a truck while updating its driving commands every 50 milliseconds. In the continuous test reported in their September 22 paper, all 616 control calculations finished before the next update was due. The test covered vehicle motion in simulation, not digging, dumping or operation at a work site.
The reverse-and-forward maneuver is difficult to automate because a wheel loader steers by bending at a joint between its front and rear halves. Its motion also changes when it shifts direction, which complicates both route planning and real-time control.
Planning the turn and predicting each direction
The team used a simplified model of the articulated vehicle to plan an initial approach to the pile. It then planned the reverse trip and forward truck approach together, choosing the stop-and-turn point as part of the route calculation rather than fixing it in advance.
To control the loader as it followed that route, the researchers trained separate forward and reverse motion models using data from an Algoryx simulation of a Komatsu WA475. Their deep bilinear Koopman models use a neural network to represent the machine's complex motion in a form that can be used for rapid prediction. A model predictive controller repeatedly forecast a short stretch of travel, issued the first steering and drive command, and recalculated from the loader's new position.
The simulated route met its deadline; model size changed solve times
In one continuous three-stage run, the simulated loader reached the team's permitted stopping region at the pile, switchback and truck. The controller solved 616 optimization problems without a solver failure. Its slowest calculation took 35.02 milliseconds against a 50-millisecond update interval.
The researchers also compared model sizes in 6,000 controller solves. Their selected compact forward and reverse models averaged 24.0 and 27.2 milliseconds per solve at a three-second prediction horizon. A larger forward model averaged 57.7 milliseconds and exceeded the 50-millisecond deadline in 89.5% of solves.
The study treated excavation separately and omitted the truck-to-start return. The paper reports one complete continuous simulated route, leaving repeated loading cycles, varying terrain and a physical loader untested in this study.
Sources
- University of California, Davis paper: Wheel-loader V-Cycle Automation with Deep Koopman MPC (September 22, 2026).
- Full paper and results tables.
- Illustrative photo of a Volvo wheel loader, not the simulated Komatsu system: Marius Vassnes, Wikimedia Commons original, CC BY-SA 4.0; unmodified.
_with_blue_excavator.jpg){kind=link}