EgoWild is a large-scale dataset of egocentric video sequences capturing dexterous, in-the-wild human interactions across diverse real-world environments. Collected using head-mounted wide-lens GoPro cameras (270° FOV), the dataset provides high-resolution recordings ideal for training embodied agents to perceive and act in natural, unstructured settings.

If you're interested in a similar dataset with the addition of motion capture gloves, feel free to check out EgoWild-X.

Unlike lab-constrained datasets, EgoWild emphasizes everyday household and assembly tasks performed without scripts or controlled conditions. Recordings include:

• Cooking: chopping, stirring, opening jars, pouring

• Cleaning & Laundry: wiping, folding, ironing, vacuuming, picking up clutter

• Assembly & Repair: building furniture, screwing parts, tightening bolts

• Daily Routines: typing, writing, opening doors, using laptops

• Utility: unpacking, carrying bags, moving furniture

Objects span from deformables (clothes, sponges, food) to rigid/articulated items (drawers, appliances, tools), with activities taking place in kitchens, garages, offices, and outdoor settings—complete with clutter, occlusion, and natural social presence.

EgoWild supports research in:

• Learning manipulation and interaction policies from egocentric visual data

• Understanding object affordances in realistic, task-driven contexts

• Advancing embodied AI systems for open-world, everyday environments

The task-rich, untrimmed nature of EgoWild makes it uniquely suited for training agents to learn from the same types of continuous, unscripted activities humans perform daily.