DATA CAPTURE
Embodied AI data capture: start from the task
Embodied AI data capture is not about stacking sensors. Define the task, valid samples, and delivery format first; then choose egocentric vision, grippers, gloves, or armbands — and plan calibration and QA.
- Embodied AI data capture
- Robot demonstration data
- Multimodal Capture
Start with the task you need to solve
Before capture, write the task definition: action segments, scene variables, failure rules, and what counts as a valid episode. Count agreed valid samples or duration — not raw filming time as trainable volume.
Labs, enterprise R&D, and scaled delivery differ in throughput and QA intensity, but share the same traceable fields: device and calibration versions, timestamps, and task metadata.
Choose modalities, not hype
Egocentric demos often use Ego Thor (Duo / Quad / Hex) by field of view; handheld teaching can use UMI-gripper; contact maps call for the tactile glove; muscle-deformation research can add FMG + IMU armbands.
Human-side observations are not robot action labels. Trajectory estimation, force mapping, and cross-embodiment transfer need extra calibration and validation.
What to check at delivery
A workable delivery usually includes a capture protocol, raw and derived packages, batch QA reports, plus version IDs, checksums, and a field dictionary. Confirm details on the commercial and engineering checklist.
Related FAQ
- How do I choose Thor-Duo, Quad, or Hex?
- Choose by field of view and task complexity: Duo for a light egocentric start, Quad for more viewpoints, Hex for richer context. Camera count is not data quality — confirm FOV, resolution, frame rate, and sync.
- What data does Thor output?
- Multi-camera vision plus IMU for egocentric ops — raw inputs for capture, calibration, and training pipelines. Keep video/IMU raw; store VIO, segments, and labels as traceable derivatives.
- What is confirmed at delivery?
- Firmware, unique SN, camera/IMU calibration files, and interface notes — plus multi-cam exposure, dropouts, battery life, and cross-device timing under real capture conditions.
Drawn from public product and solution copy. No customer logos or performance promises. Configuration and acceptance follow the project checklist.