EgoDex
A dataset of human first-person manipulation videos, used to let vision models learn human motion intent and hand-eye coordination priors.
Timeline (last 90 days)
All of REKI's pretraining data comes from public datasets, including DROID, BridgeData V2, AgiBot World, EgoDex and EgoVerse.
EgoDex's data collection requires no robot hardware participation, and data is structurally annotated at the moment of recording, requiring no human annotators or post-hoc pose estimation.
EgoDex's 194 task types cover bimanual coordinated manipulation, the domain where data is most scarce in traditional methods.
EgoDex's hand tracking accuracy can reach sub-millimeter level, far surpassing open-source solutions based on ordinary RGB cameras such as MediaPipe Hands.
EgoDex's held-out test set contains 3,243 episodes covering 111 types of tasks, which a developer parsed into MCAP format and imported into the FiftyOne visualization tool.
EgoDex's held-out test set contains 3,243 episodes covering 111 types of tasks.
Some developers have parsed the EgoDex test set into MCAP format and imported it into the FiftyOne visualization tool, with related demos and data cards published on the Hugging Face platform.
Apple's EgoDex dataset was collected by having people wear Apple Vision Pro to perform everyday tasks, with ARKit tracking 68 joint points of the hands and body in real time at a sampling frequency of 30Hz, enabling the collection of dexterous manipulation data without requiring a robot
EgoDex was accepted into ICLR 2026
A developer parsed EgoDex's held-out test set (3,243 episodes covering 111 task types) into MCAP format and imported it into the FiftyOne visualization tool
4 more timeline events
All Facts (14)
Apple's EgoDex dataset was collected by having people wear Apple Vision Pro to perform everyday tasks, with ARKit tracking 68 joint points of the hands and body in real time at a sampling frequency of 30Hz, enabling the collection of dexterous manipulation data without requiring a robot
60%UnverifiedAll of REKI's pretraining data comes from public datasets, including DROID, BridgeData V2, AgiBot World, EgoDex and EgoVerse.
50%UnverifiedEgoDex's data collection requires no robot hardware participation, and data is structurally annotated at the moment of recording, requiring no human annotators or post-hoc pose estimation.
50%UnverifiedEgoDex's 194 task types cover bimanual coordinated manipulation, the domain where data is most scarce in traditional methods.
50%UnverifiedEgoDex's held-out test set contains 3,243 episodes covering 111 types of tasks, which a developer parsed into MCAP format and imported into the FiftyOne visualization tool.
50%UnverifiedEgoDex's hand tracking accuracy can reach sub-millimeter level, far surpassing open-source solutions based on ordinary RGB cameras such as MediaPipe Hands.
50%UnverifiedEgoDex's held-out test set contains 3,243 episodes covering 111 types of tasks.
50%UnverifiedEgoDex's multimodal data contains three layers synchronized at 30Hz: egocentric video, 3D animated skeleton, and 2D pose overlay
50%UnverifiedThe EgoDex dataset contains 338K episodes, 194 task types, and 829 hours of recorded footage
50%UnverifiedApple claims EgoDex is the largest and most diverse dexterous manipulation dataset to date
50%UnverifiedEgoDex data is structurally annotated at the moment of recording, requiring no human annotators and no post-hoc pose estimation
50%UnverifiedSome developers have parsed the EgoDex test set into MCAP format and imported it into the FiftyOne visualization tool, with related demos and data cards published on the Hugging Face platform.
50%UnverifiedA developer parsed EgoDex's held-out test set (3,243 episodes covering 111 task types) into MCAP format and imported it into the FiftyOne visualization tool
50%UnverifiedEgoDex was accepted into ICLR 2026
50%