Data economics · Cost model

Cost against time, for a fixed amount of data.

This page compares the total cost to collect the same robot training data with three methods: Teleop, Primate, and UMI. A shorter collection window needs more parallel stations, so the same data costs more.

Teleop uses a robot and a paid operator per station. Primate uses a wearable gripper on a worker who is already paid for normal work. UMI uses a handheld gripper and a paid operator, but no robot.

Teleop · $0 Primate · $0 UMI · $0
39 parallel stations

The vertical line is the time window in which the data must be collected. The labeled dots show each method’s cost at that window. The × label is Teleop cost divided by Primate cost.

hours
days
$
$
$/h

The model uses three shifts per day, seven hours per shift, and 15 data points per second. One hour therefore contains 54,000 data points. Switching the data target between hours and data points changes only the unit, not the cost.

Cost formulas

Teleop

Teleop = hardware × N + 1.3 × wage × H

A paid operator works every hour. Each station uses one robot and one control desk.

Primate

Primate = device × N

The worker is already paid, so there is no wage term.

UMI

UMI = device × N + 1.3 × wage × H

A paid operator moves the gripper, but there is no robot.

Variables

H
Data target in hours. Current value: 300,000 hours.
B
Data target in billions of data points. H = B × 10^9 ÷ 54,000.
D
Collection window in days.
N
Number of parallel stations. N = H ÷ (3 × 7 × D), rounded to a whole number, minimum one. Current value: 39. One station records 21 hours per day. Teleop uses robot stations; Primate and UMI use workers with grippers.
Hardware
One Teleop robot and control desk.
Device
One Primate or UMI gripper.
Wage
One operator hour.
1.3
Wage plus 30% employer costs (taxes and benefits).

All costs are in US dollars. Hardware is bought once per station. The model does not include device life, maintenance, physical test environments, quality-control loss, downtime, inflation, or the effect of interest over time. Station counts are whole numbers, with a minimum of one.