DROID coverage sample (56 episodes)

Chosen to cover DROID's axes of variation, not sampled uniformly. A uniform draw returns mostly TRI-lab pick-and-place of rigid objects; this spreads over the properties that change what a geometry pipeline has to do.

Only camera ext1 is included. Every episode here is also covered by PointWorld's camera set, so annotations line up if you go further.

Two properties of DROID worth knowing before you look

1. current_task is usually a category template, not a description. Strings like "Move object into or out of container (ex: drawer, clothes hamper, plate, trashcan, washer)" or "Do any task, and then reset the scene." are the norm. Object identity has to come from the separate human-annotation file (aggregated-annotations-030724.json, 3 instructions per episode).
2. The human annotations cover successful episodes only. Measured over 1,153 episodes' metadata: all 832 with annotations are successes, all 125 failures have none. So any object-based filter silently selects for success. The unknown_no_annotation rows below are failure episodes drawn deliberately, and their object identity is genuinely unknown.

Object physics class

The axis that decides whether one rigid SE(3) pose per frame is even the right model.

Task template

Lab

Outcome

Episode length (frames)

min 117 / median 307 / max 2449

Files

• index.csv -- one row per episode
• manifest.json -- the picker's raw output, including all 3 human instructions
• thumbs/<uuid>.jpg -- 3-frame strip per episode (5% / 45% / 90%)
• sheets/<object_class>.jpg -- all episodes of one class stacked
• videos/<uuid>.mp4 -- the ext1 clip

Episodes