Gardening video data for physical AI
Plan useful gardening video capture with visible hand-object interactions, varied outdoor conditions and clear annotation limits.
Gardening brings together flexible plants, loose materials, hand tools and changing light. These interactions can help a research team study object handling, task order and adaptation to an irregular workspace. A useful brief identifies the particular movement or decision to observe rather than requesting outdoor footage in general.
Keep the distinction between visible activity and biological outcome clear. A video can show a seedling being placed in a pot, but it cannot establish whether the plant will thrive. Record complete, approved tasks with their initial and final arrangement so the sequence has enough context to interpret.
Tasks to include in the brief
Preparing a container
Capture arranging a pot, moving potting material and positioning the container. Specify whether the study needs the transfer itself, the order of preparation or the relationship between tools and materials.
Moving a seedling
Show the hands, plant, source container and destination together where practical. Include visible repositioning and ordinary retries without asking anyone to damage a plant or create a failure.
Watering a defined area
Describe the watering vessel, target and stopping condition. The visible sequence can support recognizing approach, pouring and withdrawal; it does not measure delivered water volume without an additional method.
Sorting tools and materials
Record choosing a small hand tool, returning it and organizing supplies. Similar-looking containers or partly hidden handles provide useful variation when their identities remain documented in the capture context.
Capture requirements
- Frame both hands and the working surface. A low camera angle may hide the inside of a pot, while a very wide view can lose the small object interaction.
- Vary container sizes, plant shapes, surfaces and ordinary daylight conditions. Record which changes were intentional so a model evaluation can separate environment variation from a different task.
- Include a short lead-in and the visible end state. A moment of repositioning can explain why a task took longer; trimming it away can remove useful behavior.
- For outcomes, distinguish a completed placement from a visibly unstable placement or an obscured ending. Do not infer root condition, moisture level or future growth from appearance alone.
Useful annotation layers
- Useful temporal labels may describe reaching, scooping, transferring, positioning and returning a tool, with the object named at an agreed level of detail.
- Preserve declared plant and material names separately from visual descriptions. Similar seedlings may be difficult to distinguish from sampled video.
- Ordinary footage does not supply calibrated soil volume, tool force, metric depth or plant-health measurements. Those require a separate collection and validation plan.
Permissions, boundaries and exclusions
- Use only activities already suitable for the participant and location. Powered equipment, chemicals and hazardous cutting are outside an ordinary gardening capture brief.
- Keep neighboring homes, bystanders, address markers and private conversations outside the recording. Permission for a garden does not automatically cover adjacent property.
- Do not repeat an unsafe movement, disturb animals or damage plants to produce a more dramatic example. Mark a missed or obscured interaction instead.