Dataset

EgoDex

A large-scale Apple research dataset and benchmark for learning dexterous tabletop manipulation.

open sourcesample readyModel hawor-cvpr-2025
Source dataset reference
Videos
390
Steps
390
Evaluation packs
4
Technical quality
84.2
Annotation quality
85.0
Hand signal
11.5
Motion density
73.5

Ranked Categories

Public scorecards

Overall QualityProvisional

92.8

362 clips · 474 votes · 14 raters

Hand-Object Interaction
92.8
Hand-Object InteractionProvisional

92.8

362 clips · 474 votes · 14 raters

Reviewed clean confidence
99.0%
Reviewed clip coverage
92.8%
Confirmed issue rate
0.0%
Confirmed clean clips
362
Confirmed issue clips
0
Needs more votes
28
Verdicts
474

Diagnostics

Non-ranked dataset signals

TrackabilityReady

85.2

200 clips · model hawor-cvpr-2025

Lost-track rate
10.0%
Relocalizations / min
1.4
IMU-only fraction
10.0%
Median feature count
130
Static pose jitter
0.40 cm
HygieneReady

82.7

200 clips · model hawor-cvpr-2025

Blur score
80.0%
Exposure score
80.0%
Dropped frame rate
0.0%
Sync offset
29 ms
Calibration completeness
80.0%
Hand SignalReady

43.8

200 clips · model hawor-cvpr-2025

Visible-hand fraction
40.0%
Mean joint confidence
50.0%
Hand temporal stability
60.0%
Interaction DensityReady

36.2

200 clips · model hawor-cvpr-2025

Hand-object co-presence
40.0%
Near-contact fraction
30.0%
Coverage / DiversityReady

60.2

200 clips · model hawor-cvpr-2025

Task-tag coverage
50.0%
Asset-title diversity
30.0%
Video coverage
100.0%
Step coverage
80.0%

Evaluation Packs

Dataset drilldowns