EmbodiedScan

Why we included this project

EmbodiedScan tackles a problem that keeps coming up in robotics and embodied AI: an agent has to read a 3D scene from its own first-person view and tie what it sees to language. The repo brings together a large ego-centric dataset, over 5,000 scans with RGB-D views, language prompts, oriented 3D boxes, and semantic occupancy, and a baseline framework, Embodied Perceptron, that fuses any number of multi-modal inputs. That pairing works as a benchmark for 3D perception and visual grounding, and the baseline models and evaluation tooling are there to adapt. It is a research resource rather than a turnkey product: the dataset requires a download application, and the code assumes a research workflow, so it suits teams working on robotics perception and embodied navigation who want a solid starting point.