Gasgoo Munich- Ontime Data — the data business arm of ON TIME— has officially launched an embodied intelligence data platform, unveiling an automated processing pipeline for ego-centric, first-person operation videos designed to train embodied AI models.
The platform targets the critical data engineering stages between video capture and training readiness. Its capabilities span core workflows: data import, AI preprocessing, action annotation, multi-tier review, automated quality checks, and standardized format export. By making raw video processing, annotation, and QC standardized, automated, and traceable, the system lowers the engineering barriers between data collection and model training.
Image credit: ON TIME
Ego-centric data is captured by head-mounted cameras, aligning the viewpoint with the operator's. Unlike third-person observation, the first-person perspective preserves authentic occlusion, field-of-view boundaries, and motion parallax — yielding a visual distribution that closely mirrors the sensory input of a robot's head camera.
NVIDIA's research in its Ego-Scale project reveals a near log-linear scaling law between the volume of ego-centric data and validation loss. That means the data isn't merely a supplement to teleoperation inputs, but an independent source of supervision offering predictable gains.
Turning raw video into a training-ready dataset, however, demands multiple steps: hand detection, camera pose estimation, action annotation, and format conversion. Relying on local scripts often leads to complex environment setups, hard-to-reproduce results, and inconsistent formats.
Ontime Data's in-house platform tackles this with an automated pipeline. It standardizes and automates the entire ego-centric data workflow — from import and annotation to verification — creating a closed loop that is both traceable and quality-controlled.
