Understand the scene
Help models recognise objects, tools, surfaces and hand-object interactions across varied real environments.
altasweep
Building a path from real-world activity to structured datasets for AI and robotics. First-person capture, privacy processing, annotation and quality review, connected from the start.

The starting point
Egocentric video records a task from the person’s point of view. It preserves the sequence of actions, the objects involved and the way hands interact with an environment.
altasweep starts with a planned Dubai cleaning pilot: professional cleaning in exchange for agreed recording of household tasks. The aim is to turn that activity into structured, useful training material.
Explore the pilotThe pipeline in development
The value is in the whole process.
Each stage makes the data more useful, consistent and ready for its intended use.
Agree the tasks, locations, recording boundaries and permitted uses before a session.
Record first-person video of real tasks. Keep session metadata and permissions connected to the footage.
Detect sensitive faces, screens and documents. Blur or mask relevant regions, then review the result.
Segment activities and label actions, objects and task stages. Connect descriptions to precise timestamps.
Review privacy coverage, label consistency and duplicates. Separate training and evaluation material by session.
Prepare versioned video, annotations, metadata and usage permissions in a documented dataset format.
Why this data matters
Structured first-person video can support several parts of an AI or robotics training programme.

Help models recognise objects, tools, surfaces and hand-object interactions across varied real environments.
Connect language to actions, learn the order of steps and evaluate whether a task stage has been completed.
Use human demonstrations to develop representations and task understanding. Robot-specific action data and validation then connect those lessons to a robot’s movement and control.
Research context: Ego-Exo4D · Open X-Embodiment
Planned expansion
Extending the same capture-to-dataset approach to industrial work, where repeatable tasks meet real-world variation.

The next direction is first-person capture around assembly, inspection, packing and material handling. Site-specific permissions, recording boundaries and sensitive production information shape each capture plan.
The goal is a wider range of demonstrations for research into task understanding, manipulation and industrial robotics.