Why robots need skilled-task data, not more footage

Robotics has its own version of the internet-scale data race. Teams are collecting first-person video of people doing everyday tasks, by the hundreds of thousands of hours, so robots can learn how hands and objects move. Volume matters. But for robots meant to do real work, volume alone isn’t enough.

Easy tasks are already well covered

Folding a towel, wiping a table, picking up a cup: these tasks are now recorded at enormous scale, often by people with no special skill, in staged rooms. That footage is valuable for general motion, and it’s getting cheaper every month.

The tasks that make a robot worth buying are different. Fastening a bracket in the right order. Pressing a seam without scorching it. Pipetting an exact volume. Packing a mixed order so nothing breaks. These are skilled tasks, and they’re scarce in today’s datasets.

Why skill changes the data

• Experts move differently. A skilled worker’s grip, order of steps and recovery from small mistakes are the very things a robot needs to learn. A beginner’s footage teaches the beginner’s habits.
• The details are the task. In skilled work, a two-millimetre misalignment or a missed check is the difference between done and redone. Only someone who does the job knows which details to show and which to mark.
• Mistakes are signal. A slip, a retry, a correction: robots learn a lot from how experts notice and fix errors. Generic footage rarely captures them, and generic labellers rarely tag them.

What skilled-task data looks like

Three things, together:

1. First-person capture of experienced workers doing real tasks, with hands, tools and parts in view, plus depth and motion sensors where needed.
2. Skilled demonstrations, including remote-controlled runs on the buyer’s robot, performed the right way, step by step.
3. Expert labelling: each episode broken into steps, with a written instruction, the tools and grasps used, and every failure and retry marked by someone who knows the work.

It has to be collected properly

Recording people at work is personal. Recent reporting on factory footage collected without clear consent has shown how easily this goes wrong. Our rules are simple: consent asked directly from each person in their own language, fair pay straight to the worker, recordings never used to track productivity, faces and bystanders blurred by default, and the right to withdraw. Buyers increasingly ask for this, and they should.

Where Kodiva fits

Kodiva vets experts for a living. We recruit people who do skilled work every day, across manufacturing, logistics, kitchens, hospitality and labs, then capture and label their work to the formats robotics teams already use.

Building a robot for a task that needs skill? Tell us the task and we’ll scope a sample dataset with you.