Sorghum has great potential to become a major feedstock for bioenergy production. For the purpose of improving its biomass yield through genetic research, manual measurement of yield component traits such as plant height, stem diameter, leaf number, leaf angle, leaf length, leaf area and plant volume in the field is the current practice, which however is extremely laborious and time-consuming. High quality 3D reconstruction of sorghum plant architecture would make it possible for automatic measurement of yield component traits. This paper presents a high-throughput field-based phenotyping robotic system capable of automatic acquisition of stereo image pairs of sorghum samples. A utility garden tractor was retrofitted with a Topcon auto guidance system to self-navigate between crop rows. A stereo camera rig consisting of six stereo camera heads was mounted in front of the robotic tractor. Stereo image pairs for each genetic line were captured at mapped locations by using the onboard RTK-GPS signals. 3D point clouds of different scenes in the field were reconstructed using PatchMatch Stereo, which effectively estimated highly slanted plant surfaces.