01
Filter the misses.
Broken capture, missing contact, wrong object, and non-transferable motion get rejected or downweighted first.
EgoArenaEgocentric Data ArenaRobot dataset quality
Robots Need Data Quality & Diversity.
EgoArena turns raw egocentric video into training signal: bad clips filtered, coverage measured, reviewers tested, consensus weighted, and policies trained on the result.
12datasets evaluating
61,080+clips reviewed
10,003hours reviewed
The quality pyramid
Raw video is cheap. Trusted signal is scarce.
A robot clip only helps training when capture is usable, the action is visible, the dataset adds coverage, and the label is trusted.
Scarce + high valueCleaner signal compounds upward
Level 6
Policy resultsrollouts prove what helped
Outcome
Level 5
Weighted training setclips ranked by signal
Highest
Level 4
Trusted labelstested reviewers, hidden checks
Trusted
Level 3
Coverage maptasks, objects, environments
Diverse
Level 2
Hard gatesvisibility, contact, capture
Filtered
Level 1
Raw robot videoabundant and noisy
Common
Common + noisy at the base. EgoArena turns useful clips into signal.
How EgoArena scores
Every clip gets a training-signal score.
02
Test the reviewers.
Golden examples calibrate humans before and during labeling.
03
Measure coverage.
We track tasks, objects, operators, environments, embodiments, and failure modes.
04
Train the policies.
Weighted data feeds models; rollouts tell us what actually helped.
Outputs
Know what is worth training on.
01
Quality report
what to keep, reject, downweight, or review
Audit
02
Coverage report
where the dataset is saturated or missing signal
Map
03
Training set
weighted clips ready for policy training
Ready
For dataset builders
Find the clips worth training on.
Submit raw egocentric data. EgoArena turns noisy clips into ranked, reviewable signal.