Development of a Virtual Data Collection System for VLA Learning of Agricultural Humanoid Models

Abstract

High-quality data collection is indispensable for Vision-Language-Action (VLA) learning in humanoids to address labor shortages in agriculture. Conventional robot simulators often lack visual fidelity for complex natural environments. This study constructs a data collection system that integrates a photorealistic environment using Unreal Engine 5 and 3D Gaussian Splatting with intuitive VR operations. By performing demonstrations in a virtual space that mirrors the actual robot's kinematics, we aim to generate data that captures the operator's intent. This report describes the system architecture and the implementation of a 29-DOF robot model based on official kinematic specifications.