How this guide is researched and how the cost model works

Contents

1. Evidence we accept
2. How the cost model works
3. What the model leaves out
4. The featured service
5. Keeping it current

A how-to guide is only as good as the numbers under it. This page sets out what we accept as evidence, how the cost-per-clean-episode model on the main guide is assembled and what it cannot tell you.

Evidence we accept

1. Research papers and their project pages, for what a rig costs to build, how many demonstrations a task needed and how long a dataset took to collect.
2. Official documentation, for data formats and tooling.
3. Makers' and vendors' own published pages, for prices, read on the access date shown. Figures a company reports about itself are labeled as vendor-reported.
4. Official statistics and dated job listings or press reports, for what operators are paid and what employment adds to a wage.

We do not use anonymous testimonials or figures we could not open and read ourselves. When a number is not published, the page says "not published" and does not estimate it.

How the cost model works

The model on the main guide divides what an hour of collection costs by how many usable episodes that hour produces:

cost per clean episode = (operator cost per hour + rig cost per hour) ÷ (episodes recorded per hour × share that pass quality checks)

• Operator cost per hour is a published wage, converted to an employer cost with official compensation statistics, or a published hourly price for a hired operator.
• Rig cost per hour is a published hardware price spread over a stated number of collection hours.
• Episodes per hour and pass rate are the two inputs no source can supply for another team's task. The table states the values it assumes and shows how the result moves when they change.

The model is an illustration of how the inputs combine. It is not a quote, and it is not a forecast of any provider's price. The only way to get real values for the last two inputs is to measure them on the reader's own robot and task.

What the model leaves out

Engineering time to build the recording pipeline, recruiting and training time, supervision, facility costs, robot downtime and repairs, and the cost of storing and labeling the data. The in-house or hired guide discusses each.

The featured service

The sidebar features RoboPilots, a managed operator service, because the guide's last step, deciding who operates the robot, is the one such a service addresses. Its entry is built only from its own published pages, with its limits stated.

Keeping it current

Sources show the date we accessed them. Priced inputs are re-checked at least once a quarter, and each page logs material changes.

Further reading

• Guide · Throughput and qualityEpisodes per hour and quality checks in teleoperation data collectionTwo numbers decide what a dataset costs: how many episodes an operator records in an hour and how many of them pass. What published collections report on both.
• Guide · StaffingIn-house or hired teleoperators for data collection: the costs comparedOnce the rig works, the operator is most of the cost. The published pay ranges, the true employer cost, the published prices for hired collection and a way to compare them.
• Guide · FormatsRobot data formats compared: LeRobot, RLDS, HDF5 and MCAPThe format decides whether the training job can read what the operators recorded. What each of the four common formats stores, from its own documentation.