AIAny
Icon for item

HubEgo-100K

Provides first-person video of skilled manual work with synchronized motion-sensor data for robot learning and dexterous manipulation research. The first release spans 10,124 hours from 12,227 contributors, with task-level annotations.

Introduction

Robot-learning datasets often trade scale for control: lab demonstrations are precise but narrow, while ordinary first-person video is broad but difficult to align with physical motion. This collection takes the latter route while adding synchronized IMU streams, offering a large view of hands working across real trades and environments. Its breadth is useful for studying visual generalization, but it is not a ready-made set of robot action trajectories.

What Sets It Apart
  • The first release contains 10,124 hours across 72,754 episodes and 15,548 environments, recorded by 12,227 contributors. This variety can expose models to tools, materials, and work settings rarely represented in curated lab demonstrations.
  • Hands appear in more than 95% of footage, and 96.9% of retained time is on-task. That concentration makes the videos more relevant to manipulation learning than general egocentric activity footage.
  • Frame-aligned accelerometer and gyroscope readings complement the video, while structured task and subtask labels support temporal and language-conditioned analysis. The annotations were generated with a vision-language model and have not been fully checked by people, so they should be treated as noisy labels.
  • The footage captures human work rather than robot control signals or standardized hand poses. It can inform visual and semantic learning, but does not by itself provide the action supervision needed to directly train a robot policy.
Who It’s For—and the Trade-off

A good fit for researchers exploring large-scale egocentric perception, task understanding, or pretraining for embodied systems, especially where real-world diversity matters more than controlled capture. Look elsewhere if you need verified annotations, robot joint/action trajectories, or consistent hand-pose labels. The CC BY 4.0 license permits commercial use with attribution; access requires a short request form, and privacy filtering removes sensitive footage.

Information

  • Websitehuggingface.co
  • OrganizationsHub Labs (Hub Data, Inc.)
  • Published date2026/11/12

More Items

Hugging Face

Evaluates whether GUI agents can locate interface elements from Korean instructions and website screenshots. Its 3,319 human-annotated examples span 142 websites, providing a focused test of visual grounding on Korean web interfaces.

Hugging Face

Evaluates document parsers across text, formatting, tables, charts, layout grounding, and forms. Severity-aware scoring prevents a small number of critical errors from being hidden by otherwise high page-level accuracy.

Hugging Face

Provides 220 Korean and English GUI/web-agent tasks with evidence checks to assess task completion, observed evidence, and whether final-answer claims are grounded in that evidence.