Examining the simulation-to-reality gap of a wheel loader digging in deformable terrain

TL;DR

This study quantifies the sim-to-real gap for a full-system wheel loader operating in deformable soil by comparing field measurements with eight fidelity levels of DEM and multiscale terrain simulators. The results show a typical sim-to-real error around $10\%$, with the reduced multiscale G models delivering near real-time performance and sometimes outperforming full DEM in realism. A force-based controller trained in a fast simulator transfers with modest loss to higher-fidelity domains, indicating limited domain sensitivity and useful predictivity for calibration. The findings imply that most discrepancies stem from model errors rather than numerical issues, and that co-simulation-based multiscale terrain models can achieve substantial speedups without a large accuracy penalty. Practical implications include guiding simulator design and calibration strategies to enable reliable training data and transferable autonomous control for earthmoving tasks.

Abstract

We investigate how well a physics-based simulator can replicate a real wheel loader performing bucket filling in a pile of soil. The comparison is made using field test time series of the vehicle motion and actuation forces, loaded mass, and total work. The vehicle was modeled as a rigid multibody system with frictional contacts, driveline, and linear actuators. For the soil, we tested discrete element models of different resolutions, with and without multiscale acceleration. The spatio-temporal resolution ranged between 50-400 mm and 2-500 ms, and the computational speed was between 1/10,000 to 5 times faster than real-time. The simulation-to-reality gap was found to be around 10% and exhibited a weak dependence on the level of fidelity, e.g., compatible with real-time simulation. Furthermore, the sensitivity of an optimized force feedback controller under transfer between different simulation domains was investigated. The domain bias was observed to cause a performance reduction of 5% despite the domain gap being about 15%.

Figures (15)

• Figure 1: Photo from the field test.
• Figure 2: Illustration of the wheel loader from the side and the top with the quantities measured during the field test marked in red, joints in blue, and actuators in green.
• Figure 3: Images from the eight simulators of different levels of fidelity. In the type D-simulators (left column), the gravel pile is fully resolved in particles with characteristic diameters of 50, 100, 200, and 400 mm (top to bottom). In the type G-simulators (right column), a multiscale technique is applied with different grid sizes 50, 100, 200, and 400 mm (top to bottom).
• Figure 4: Illustration of the multiscale terrain model (left) adapted from Servin2021. It can be regarded as a heavily reduced version of the full DEM model (right). The region of active soil movement is predicted and substituted with a rigid aggregate body that couples with the vehicle dynamics (upper left). The mass flow inside the active region is co-simulated using a DEM model (lower left).
• Figure 5: Simulation of the HD27 test with overlaid images from the D50 and G200 simulators at one-second intervals starting from time $0.37$ s. The D50 particles are color-coded by speed, with blue for 0 and red for 1 m/s, while the G200 particles are gray. The shape of the active zone and the distributions of mass in the bucket are in good agreement.