Physics Validation of Simulation Engines for Plant-Robot Interaction in Agricultural Settings

The development, training, and testing of agricultural robots requires simulation environments that combine photorealistic sensing with physically credible plant interaction. While modern game engines offer high visual fidelity, their suitability for plant-level dynamic interaction remains insufficiently validated for robotic applications. This study presents a hybrid UE5-ROS2 simulation framework and systematically evaluates the physical behavior of Unreal Engine 5 (Chaos) against Gazebo Classic (ODE) for articulated plant models. A unified cross-engine architecture enables deterministic model reconstruction and synchronized state comparison under identical boundary conditions. Verification and validation are conducted in two stages: canonical benchmark systems (mass-spring-damper, pendulum, and bouncing ball) are used to verify solver-level numerical behavior, followed by plant-level cross-engine validation experiments on articulated strawberry models. Results show that although the Chaos solver exhibits stronger cumulative numerical dissipation compared to ODE, both engines produce stable and bounded dynamic responses within short-horizon interaction regimes. These findings support the use of UE5 as a credible simulation platform for perception-driven agricultural robotics when long-term elasticity fidelity is not the primary objective.