From Restocking Shelves to Robot Skills: What Ego2Robot Shows About Everyday Human Data

A store associate rolls a cart into a stockroom, scans a package, checks a shelf location, turns the item to face forward, and moves an older product ahead before placing the new one behind it. To a person, this is ordinary work. To a robot-learning system, it is a compact lesson in visual search, sequencing, grasping, orientation, placement, and exception handling.

Released on August 3, 2026, Ego2Robot explores how first-person human manipulation video can be converted into robot-format training data. The research pipeline uses action retargeting, robot-arm visual synthesis, and several stages of quality curation. It reports 18,561 hours of synthesized training data across 15 robot morphologies, produced from approximately 1,940 hours of source egocentric video.

The research matters because useful demonstrations are everywhere outside robotics labs. Retail employees restock shelves and prepare displays. Warehouse teams sort parcels and fill totes. Hotel staff arrange rooms and replenish supplies. Small-workshop employees assemble, inspect, and pack products. These familiar tasks contain repeatable hand-object interactions that can be recorded without placing a robot at every location.

Turning those videos into robot data is not automatic. Human hands and robot end effectors have different shapes, reach limits, and movement constraints. Ego2Robot aligns human motion with robot actions, changes the visual embodiment, and filters unsuitable outputs. Its experiments report improved generalization when synthesized data is combined with robot demonstrations, especially when appearance, embodiment, or task conditions change.

The quality of the original recording still sets the ceiling. A system cannot reliably reconstruct a product pick if the hand leaves the frame, a label is unreadable, or a fast head turn creates severe distortion. Everyday collection therefore needs visible hands and objects, stable task boundaries, natural but well-framed movement, and examples of ordinary variation such as different shelf heights, package shapes, carts, lighting, and worker approaches.

EGO R9 can support that upstream capture in practical workplaces. Its hands-free head-mounted viewpoint follows the wearer through scanning, sorting, stocking, packing, and inspection. A 120-degree wide-angle view helps keep both hands and nearby shelves in context, while 1080P global-shutter video is suited to movement-heavy work. Its 6-axis IMU above 200 Hz, shared clock, and global timestamps can help align video with task markers or approved external systems.

A responsible pilot should begin with a small number of routine, low-risk workflows. Define what starts and ends each task, which areas may be recorded, how customer or employee information will be excluded, and what makes a clip acceptable. Participation, retention, access, and intended model use should be explained clearly, and the camera must never override safe lifting, required breaks, or normal work procedures.

Ego2Robot points toward a future where ordinary work can contribute to more adaptable robot policies after careful alignment and curation. The EGO R9 does not perform that synthesis by itself; it provides the consistent first-person visual and motion evidence on which downstream conversion can build. For data teams, that makes familiar daily tasks a valuable starting point rather than background activity.