One configurable unit. The same kit at every site.
The same hardware and software at every site, so deployment is fast and the data is consistent everywhere. Teams on site get a working deployment and a dashboard. Every run becomes a clean, structured episode.
An RGB-plus-depth rig over the task cell, mobile or fixed, with no rewiring. Off-the-shelf, e.g. Orbbec Gemini on NVIDIA Jetson.
A feed from the robot itself: joint encoders, force-torque and end-effector state, through ROS2 or the vendor API. Brand-neutral.
A teleop rig and optional wearable so people seed demonstrations and correct failures. The corrections are the most valuable data we collect.
An on-site Jetson-class box time-syncs every stream and blurs faces and badges before anything leaves the building.
Defines the task, splits it into episodes, and tags success or failure.
Labels objects, actions and outcomes with vision-language models, using the robot’s own telemetry as ground truth.
Stamps every episode with source, time and consent basis, and anonymizes to the entity level.
Uptime, throughput, and what the robot can and cannot yet do, in real time.
Every episode carries its lineage from the moment of capture to the moment of delivery.
Every episode arrives cleared, traceable, and safe to train on.
Data is captured under agreement and delivered cleared for AI training and evaluation. No scraped sources, no murky provenance, nothing you have to take on faith.
Faces and badges are blurred on site, before anything leaves the building. No personal identifiers are retained, so the corpus is non-personal by construction.
Every record carries full lineage and a consent basis. We warrant provenance and indemnify on delivery, so your data governance is covered from day one.
Documented governance and provenance make the models you train easier to certify, not harder. That is the difference between data you can ship on and data you can only experiment with.