First-person video of real household chores, capturing the dexterous hand movements robots need to learn manipulation. License it off the shelf as raw video, annotated with our standard schema, or with fully custom annotation.
EgoChores is a targeted, high-density egocentric video dataset of everyday household cleaning and tidying, captured from a head-mounted first-person perspective with a focus on dexterous hand-object manipulation. Our goal was to address something we found lacking in currently available datasets: a single, high-demand task family, available off the shelf, with production-grade annotations optimized specifically for pre-training embodied AI.
Precisely annotated in Label Studio, every clip is segmented into timed actions and resolved against a hierarchical object taxonomy created under the supervision of robotics researchers. Commission custom collection runs in the same schema to extend and adapt the dataset.
EgoChores ships with a fully published, machine-readable annotation schema. Your robotics teams can train on it natively and later use the exact same taxonomy to commission custom collection runs.
Rich, textual step descriptions aligned precisely to visual frame boundaries.
Continuous first-person trajectories tracking multi-modal hand states and object interactions.
Clear pre-condition and post-condition states captured before and after every major task boundary.
High-density, domain-specific data to bridge the simulation-to-reality (Sim2Real) gap.
License the dataset at the level of annotation your team needs — or request a free sample to evaluate it first.
The full EgoChores video corpus with basic clip-level tags and metadata.
Teams running their own annotation pipeline, or pre-training directly on raw egocentric video.
EgoChores video with the HumanSignal standard label schema applied — action segmentation, dexterous hand-object interaction, and object taxonomy.
Immediate integration into VLA training loops and manipulation benchmarking.
EgoChores video with heavier, bespoke annotation — custom object classes, edge cases, and proprietary schema configurations.
Production-grade robotics labs scaling toward deployment.
Capture methodology, taxonomy design, and benchmark results.
The machine-readable schema your data loaders train on natively.
Versioned splits, dataset card, and download instructions.
While Ego4D provides unparalleled unconstrained scale, it lacks the dense task-family specificity needed for highly repeatable robotic policy learning. EgoChores provides tight, standardized taxonomies across identical task categories, minimizing background noise and maximizing target interaction density.
Yes. Our human data collection networks span multiple global regions, allowing us to capture diverse architectural styles, appliance form factors, and cultural variations in chore performance.
EgoChores ships natively in standard JSON / COCO-style formats for vision models. Our internal annotation tooling also supports custom schema compilation to format outputs directly for your team's proprietary data loaders.
Let's design your proprietary dataset program today.