Distributed multi-robot mapping cuts communication by orders of magnitude

A team of researchers has introduced a new distributed framework for multi-agent Next-Best-View (NBV) selection that dramatically reduces communication overhead while preserving mapping quality and safety. The system, outlined in a paper submitted to IROS 2026, tackles the challenge of safe path planning in unknown environments by having each robot maintain a private local 3D Gaussian Splatting (3DGS) map. Instead of sharing raw sensor data, robots exchange only candidate viewpoints, planned trajectory descriptors, and scalar expected information gain contributions. Coordination is achieved via Consensus ADMM (C-ADMM) over a communication graph, ensuring the team jointly maximizes expected information gain within masked zones along trajectories. Collision risk is modeled using Average Value-at-Risk (AV@R) computed from each robot's local 3DGS map, which both shapes the masking radius and scores planned paths.

The framework was evaluated in Gibson environments across multiple team sizes. Results show that the distributed formulation approaches the centralized baseline in terms of mapping completeness and trajectory safety, but with communication reduced by orders of magnitude. This makes the approach highly scalable for large robot teams operating in bandwidth-constrained or communication-denied environments. By localizing sensor processing and limiting exchanges to high-level planning artifacts, the method enables robust multi-agent exploration without the overhead of raw data sharing. The use of 3DGS as a compact scene representation is key to balancing information richness with computational efficiency. This work represents a significant step toward practical, risk-aware multi-robot systems for applications like search-and-rescue, planetary exploration, and industrial inspection.