dreamlake-hug

A 15-recording demo slice of kevinywu/1m-hugs (HUG: Human Universal Grasping, Lerrel Pinto's lab, NYU — CC-BY-4.0, arXiv:2606.17054, grasping.io), repacked for the DreamLake dataset-viz folder format.

The source ships tar shards of Python pickles; every payload inside is re-emitted here as an established format — JPEG / 16-bit PNG frame sequences, COCO keypoints, a Parquet point-track — with nothing else changed. One episode per grasped object, from train shard 0. Built in-cloud by packages/viz/scripts/build-hug-demo.py (viz-workspace); probe facts in scripts/hug-probe.py.

What the hand is. No hand is visible in any frame: each recording is the egocentric approach toward an object, and every frame carries the same eventual grasp as its annotation (one grasp_index per recording) — HUG's task is predicting the grasp from the scene, so the MANO skeleton is the label, not a detection. The COCO keypoints are that grasp reprojected into each frame; the 3D pose track is the same label in per-frame camera coordinates.