Published on arXiv (https://arxiv.org/abs/2603.23676), the paper addresses long-horizon planning in 3D environments using only visual observations and under-specified natural-language goals. The authors identify two shortcomings in existing approaches: symbolic planners with brittle relational grounding, and direct action-sequence generation from 2D vision-language models (VLMs), both of which struggle with many-object reasoning, rich 3D geometry, and implicit semantic constraints. RAMP-3D processes RGB-D observations alongside natural-language task specifications to reactively generate multi-step pick-and-place actions. The system predicts paired 3D masks — a 'which-object' mask identifying what to pick and a 'which-target-region' mask specifying placement location — extending existing 3D grounding models that ground natural-language referents to 3D segmentation masks. Experiments were conducted across 11 task variants in warehouse-style environments containing between 1 and 30 boxes with diverse natural-language constraints. RAMP-3D achieved a 79.5% success rate on long-horizon rearrangement tasks, significantly outperforming 2D VLM-based baselines, with the authors concluding that mask-based reactive policies represent a promising alternative to symbolic pipelines for long-horizon planning.