Translation averaging seeks to recover global camera locations from noisy pairwise translation directions and is a central component of Structure-from-Motion. The problem is challenging because direction measurements contain no distance information. As a result, standard formulations can be ill-conditioned, collapse many cameras to a single point, or be misled by structured outliers that support a coherent but incorrect reconstruction. I will present TriP, a local-to-global framework based on a simple geometric observation: the directions along a nondegenerate triangle determine the relative lengths of its three edges. TriP treats each triangle as a local reconstruction piece, known only up to an unknown scale, and stitches these pieces together by requiring overlapping triangles to agree on their shared edges. In the logarithmic domain, this becomes a robust one-dimensional synchronization problem on a graph whose nodes represent triangles. The synchronized edge-length estimates are then used to recover the global camera locations. This intermediate metric layer has two important benefits. It rules out zero-scale collapse by construction and provides higher-order redundancy against adversarial, cycle-consistent, and other structured corruptions. I will discuss deterministic exact-recovery guarantees under bounded adversarial corruption, including a nonvanishing corruption tolerance on complete observation graphs. I will also present experiments on synthetic and real Structure-from-Motion datasets, where TriP improves robustness and accuracy over existing translation averaging methods while scaling to graphs with more than one million cameras. Joint work with Zhekai Fan, Wanze Li, and Jinxin Wang.