UniT: Unified Geometry Learning with Group Autoregressive Transformer

UniT introduces a unified feed-forward transformer for 3D geometry perception, covering both online and offline inference.

It integrates multiple input modalities, uses queue-style KV caching for long-range memory, and adds a scale-adaptive loss for better metric-scale handling.

The model aims to replace several separate 3D geometry pipelines with one scalable system for point maps and spatial understanding.

Reported benchmark results are strong across diverse tasks, suggesting simpler deployment for robotics and vision systems.