Ground-penetrating radar is a mature technology which has promise as a solution for humanitarian demining. The technology is fast, inexpensive, and capable of detecting both metallic and non-metallic landmine casings. However, the rough air-ground interface below which anti-personnel mines are buried, reduces the efficacy of air-coupled GPR by increasing clutter and masking target responses. Recent literature focuses on optimizing signal processing techniques to remove the effects of the surface and reliably extract the target reflection. Conversely, this work proposes the use of ground-contact antennas, which greatly improve signal penetration and are less affected by ground clutter, thereby simplifying data analysis. Achieving contact between the surface and the antennas is done by integrating the antennas onto the feet of the Walking Tri-Sphere, a non-articulated walking robotic platform designed by Square One Systems Design (Jackson, WY, USA). Rather than imaging the subsurface, localization of potential targets is achieved using a robust geometric analysis, minimizing the required number of GPR scans. Overall, by using fewer scans and simpler data processing techniques, this method is capable of increasing the surveying speed of traditional GPR methods. The proposed detection system is evaluated experimentally using the P400 ultra-wideband impulse radar from Time Domain (Huntsville, AL, USA), and computationally using a 3D finite-difference time-domain model. Compact spiral antennas which operate from 3-6GHz were designed considering the application and desired coupling into the ground. The polarization and directivity of the antennas minimizes the direct signal, simplifying the identification of target reflections. Subsurface scans which satisfy both an amplitude and correlation threshold are then analyzed with a localization algorithm, which utilizes time-difference of arrivals to geometrically determine the target location. A minimum of four unique bistatic GPR scans are necessary to evaluate for the target's position, and an increased number of GPR scans improves the accuracy and reliability of the results. Using the proposed localization method, metallic cylindrical targets are successfully located experimentally. Consideration of non-metallic targets is also addressed experimentally and more extensively computationally. Overall, the proposed method provides a viable solution for autonomous pre-screening of an area for humanitarian demining.