Overlapping Domain Decomposition for Distributed Pose Graph Optimization

A team from MIT led by Aneesa Sonawalla, Yulun Tian, and Jonathan P. How has introduced ROBO (Riemannian Overlapping Block Optimization), a novel distributed algorithm for multi-robot pose graph optimization (PGO) accepted to ICRA 2026. PGO is the core mathematical problem behind simultaneous localization and mapping (SLAM), where robots must collaboratively build a consistent map of an environment. ROBO innovates by employing overlapping domain decomposition, a technique that allows neighboring robots to share a controlled subset of their estimated poses (positions and orientations). This creates overlapping optimization blocks within the global pose graph, fundamentally altering the convergence dynamics compared to fully distributed methods where robots only share minimal boundary data.

The key technical achievement is ROBO's tunable trade-off between communication overhead and convergence speed. By sharing an average of just 36 Kilobytes of additional pose data per iteration between robots, the algorithm achieves a 3.1x faster convergence in terms of iterations compared to leading distributed PGO baselines. The researchers also developed an asynchronous variant robust to network delays, making it suitable for real-world deployments where communication is unreliable. This work provides roboticists with a practical, flexible tool for scalable multi-robot systems, enabling faster, more accurate collaborative mapping in search-and-rescue, exploration, or warehouse automation scenarios where communication bandwidth is a limiting factor.