Recording useful video for physical AI

Choose a familiar activity with an observable result

Useful physical-world footage shows what a person did, what they interacted with and what changed. Choose work you already understand: repotting a plant, folding a towel, sorting components, wiping a worktop or packing an unlabelled box. A clear, repeatable action is easier to describe and review than a long recording in which the main task is unclear.

Record ordinary safe practice in a location where recording is authorized. Do not create a hazardous situation to make a clip more interesting. A failed attempt can still be informative when it happens naturally and is described honestly. The point is to preserve observable behavior, not to manufacture a perfect performance.

Show the beginning, the action and the result

Start before the relevant interaction so the viewer can see the objects and initial state. Keep recording through the action and long enough to reveal the result. For potting a plant, show the empty pot, the addition of soil, the placement of the plant and its final state. Beginning halfway through a motion removes context a model or reviewer may need.

Avoid decorative edits, speed changes and transitions in the source recording. These can obscure timing or create false continuity. If an unavoidable interruption occurs, describe it. A file that joins several takes should not be presented as one uninterrupted action, and a repeated segment should not be counted as additional unique activity.

Choose the viewpoint deliberately

A head-mounted camera follows the wearer’s perspective and often captures the relationship between attention and hand activity. It can also move quickly or point away from the work. A fixed camera provides a steadier view of a work surface, but a body or object may block the interaction. Neither viewpoint is automatically better for every research objective.

Make a short test recording before a long session. Check whether the hands, tools and target object remain visible at important moments. Adjust the mount or camera position, then record the actual setup as head-mounted, chest-mounted, handheld or fixed. Do not call a fixed overhead view egocentric merely because the person performing the action uploaded it.

Preserve image detail and report actual settings

Use stable lighting and enough focus to distinguish the manipulated objects. Avoid strong backlighting, a covered lens and unnecessary camera movement. Keep the original file when possible. Repeated messaging-app exports or aggressive compression can remove small details such as contact points, fast movements and the state of a tool.

Report the actual resolution, recording device and frame-rate setting. A declared frame rate is not an independently measured one, and a file may use variable timing. Do not enter a higher number because it seems more valuable. If audio is unnecessary and creates privacy risk, plan the capture accordingly; otherwise review it with the same care as the image.

Describe the task in terms another person can check

Select the activity category that best matches the recording, then add a specific task, environment, objects and tools. Gardening is a category; filling a plant pot with soil is a task. A greenhouse workbench is an environment; a scoop is a tool. These distinctions help a buyer find relevant footage without relying on a vague title.

For several professions or activities, maintain separate relevant listings rather than forcing everything into one description. Record important limitations, such as the final result being hidden or the first step occurring before the camera started. Useful metadata makes uncertainty visible. It should not claim expertise, location diversity or measured sensor properties that the recording does not establish.

Add temporal annotations that stay close to the evidence

A basic action annotation identifies when an interaction begins and ends, the verb, the object, the intended goal and the visible outcome. Use time relative to the start of the uploaded file. In the normalized record, 1.5 seconds becomes 1,500 milliseconds. Keep each interval within the recording and ensure its end follows its start.

Describe what is visible: place a seedling into a pot is more useful than do gardening. Mark success only when the result can be observed; otherwise choose uncertain. Ego4D’s annotation documentation separates narrations, temporal events and spatial labels, while Ego-Exo4D documents timed task steps. These are different annotation products, not interchangeable names for a caption.

Contributors can supply task knowledge and a first layer of action timing. Dense object tracks, segmentation masks, hand keypoints or calibrated 3D labels may require trained annotators, dedicated tools and independent quality checks. Add these only against a defined buyer specification. Automatically generated labels should retain their origin and review state.

Complete the rights and privacy check before uploading

Confirm that you can supply the recording, that relevant participants have the necessary information and permissions, that the workplace or property allows this use, and that private information has been removed. Do not film customer addresses, live accounts, confidential drawings or someone else’s conversation just because they happen to be near the task.

Watch the whole file, including reflections and the edges of the frame, and listen to the audio. A privacy edit needs another review after export. Ego4D’s published process includes participant control and human review alongside technical de-identification. This is a useful reminder that a blur tool or automatic caption cannot certify that every privacy issue has been resolved.

Understand what the recording can and cannot provide

Human video can show physical interactions, task structure and changes in objects. It does not directly contain a robot’s joint positions, gripper commands or force measurements. DROID, for example, documents robot-specific observations and actions collected with its setup. Converting a human video into another file format cannot recover those unrecorded signals.

Upload completion is also separate from review, approval and a buyer’s purchase. Review may identify a poor view, inconsistent timing or a permission concern that needs correction. Preserve your source and the context needed to explain it. A carefully framed, accurately described recording gives a laboratory a stronger basis for judging suitability than a larger volume whose limitations are hidden.