Anna-Maria

Identifies a shortcut in how 3D vision-language models learn grounding: when asked about an object marked by a bounding box, the box's size and position already give away much of the answer, through both the text prompt and the model architecture. Replaces the box with a single 3D point, under which existing models lose much of their accuracy, and introduces P³-VLM, a 3D VLM designed for point prompts.

Introduces the first 3D vision-language model operating natively on Gaussian splats, without object detectors. A dual sparsifier distills about 40k language-aligned Gaussians into 132 task-aware tokens, achieving state-of-the-art results across 3D VQA benchmarks and fivefold gains out of domain.

Introduces a dataset of 280 indoor scenes in Universal Scene Description format, annotated with object and part segmentation, motion parameters, movable and graspable regions, connectivity, and object mass, so scenes can be used directly in downstream tools. Proposes USDNet, which jointly segments movable and graspable parts and predicts their articulation from scene point clouds, and establishes a benchmark for articulated scene understanding.