OSMa-Bench++: Toward Open-Ended Benchmarking of Semantic Mapping for Manipulation with Prompt-Generated Synthetic Scenes

Researchers introduced OSMa-Bench++ to evaluate semantic mapping for robot manipulation in a more open-ended way.

The pipeline uses SceneSmith to generate scene descriptions, synthesize matching indoor environments, and convert them into benchmark-ready simulations.

It also adds prompt-grounded semantic questions to stress-test mapping methods across clutter, small objects, occlusions, and lighting changes.

The framework aims to make semantic map evaluation more realistic and better aligned with manipulation challenges.