SCOUT framework lets robots actively explore and understand scenes

SCOUT (Semantic scene COverage via Uncertainty-guided Traversal) is a new framework from Junyu Mao, Sara Ayoubi, and colleagues that closes the loop between robot exploration and scene understanding. Traditional pipelines treat perception as a post-processing step over fixed data, but SCOUT operates online: it incrementally builds a 3D scene graph where nodes maintain fused geometry and posterior beliefs over open-vocabulary object labels, and edges encode structural relations like "on", "inside", "belongs to", and "next to". These beliefs feed into an uncertainty-guided traversal planner that selects next viewpoints by balancing expected semantic certainty gain, geometric coverage gain, and travel cost.

This approach means a robot can decide to revisit an ambiguous object when additional evidence would resolve uncertainty, or expand into unseen free space when the scene is incomplete. The system treats semantic scene completeness as an operational objective, not a passive byproduct. Accepted at the 2026 ICRA Workshop on Uncertainty in Open World Robotics, SCOUT moves toward autonomous agents that can patrol, update, and reason about evolving indoor environments with minimal human intervention. The code and data are expected to be released alongside the paper.