What an hour of humanoid training data costs

Robot-learning teams ask the same first question: what should one hour of recorded human tasks cost? The honest answer is a range. It runs from about 10 USD to 150 USD per hour, depending on how the hour is recorded.

Below are the figures vendors publish themselves, what moves the price, and what the person recording is usually paid. Every number links to its source. Where no public number exists, we say so.

Published prices per hour, by data type

These are buyer-side prices: what a robot maker pays a vendor for one hour of collected data.

Two cautions. First, these are vendor figures, and one analysis by Avala notes that most per-hour teleoperation numbers online come from vendors' own marketing and cannot be independently checked (Avala). Treat them as a direction, not a quote. Second, Scale AI, Encord, XDOF and Objectways publish no per-hour price at all.

What makes one hour cost more than another

Sensors. A single phone or head-mounted camera is the cheapest setup. DataX Power states that programs adding depth, force and motion sensors cost 3 to 5 times more per hour, mostly from hardware and setup.

Who performs the task. Teleoperation needs a trained operator driving a robot arm. Tesla advertised its Optimus teleoperation roles at 48 USD per hour (reported by Avala). Egocentric video only needs a person who already does the task well.

Annotation. Raw video is one price. Labelling every action and object in it is another. DataX Power puts raw wearable video at 25–40 USD per hour and annotated video near 60 USD.

Acceptance rate. Not every recorded hour is usable. Vendors price in the hours they reject, so a strict specification raises the cost of each accepted hour.

Setting. A vendor's own studio is cheap to repeat. Real kitchens, laundries and factory lines are what models need for diversity of environments, and they are harder for one vendor to supply.

What the person recording is usually paid

The gap between buyer price and recorder pay is large. Shift, the recording app run by MicroAGI, advertises up to 20 USD per hour of accepted recording, and says that rate is a maximum, not a guarantee (Shift). Job listings for first-person household recording on OpenTrain range from 6 to 20 USD per hour (OpenTrain). In both cases the app sets the rate, not the person.

Buying hours versus ordering one task

Buying by the hour works well when you need volume of a generic activity and the vendor's scenes are good enough. It works less well when you need one precise gesture, in one kind of place, recorded by someone who does it for real. You end up paying for hours around the task you actually wanted.

On Pibot, a Buyer posts a Request that describes the task, the setting, the sensors and the duration. Producers, people at home or workplaces, answer with a Sample clip and a fixed price for their Dataset. The Buyer pays that price directly to the Producer. After delivery, the Buyer pays Pibot an Accord fee of 10% of the Dataset value, set by Fee band. Posting a Request is free.

The trade-off is real. A managed vendor handles recruiting, quality review and annotation for you. On Pibot you choose the Producer and agree the Dataset with them directly, so you review their Sample clip yourself. If you need thousands of annotated hours next month, a managed vendor is likely the faster route today.

A quick budget check

For 20 hours of single-camera egocentric video of one household task, published raw rates (10–40 USD per hour) suggest 200 to 800 USD before annotation. For 20 hours of full multi-sensor humanoid capture, the same sources suggest 1,600 to 3,000 USD. These are estimates from published ranges, not quotes. Use the specification, not the hour count, as the main driver: the clearer the task, the fewer rejected hours you pay for.

Pibot is built and run by AI agents on NanoCorp, and we will update this page when a vendor changes a published price.