Robots Have 4,000 Public Hours of Training Data and Need an Estimated 1 Billion, at $10 to $100 Each

The largest public robot dataset holds about 4,000 hours of demonstrations, and the estimate for what a general-purpose robot needs is about 1 billion hours, a gap of 250,000 times that would cost between $10 billion and $100 billion to close at today's collection prices. Both figures come from Brian Potter's October 15 essay How to Train Your Robot on Construction Physics, and the second is his estimate, not a measurement. The arithmetic around it is worth doing slowly, because the number that matters is not the dataset size but the price per hour of usable data.

What the public datasets hold

Open X-Embodiment, assembled by Google DeepMind and academic labs, pools 60 existing datasets covering 22 robot embodiments, with more than 1 million trajectories, 527 skills and 160,266 tasks, according to the project page. Its authors report that RT-1-X beat RT-1 trained on individual datasets by 50% in the small-data domain, and that RT-2-X beat RT-2 by 3x on emergent-skill evaluations. Potter puts the dataset at about 4,000 hours.

Physical Intelligence's Pi 0 added its own dexterous dataset from 8 robots to Open X-Embodiment and internet vision-language data, and was published on October 31, 2024. The company's post gives no hour count; Potter's figure of 10,000 hours is his own, so treat it as an outside estimate. The post does say "There is no such treasure trove of robot data" as there is for text.

1. Open X-Embodiment4000 hours
2. Pi 0 (Potter's estimate)10K hours
3. Generalist GEN 1 (UMI data)500K hours
4. Dyna Robotics (claimed, egocentric video)1M hours

Source: Brian Potter, Construction Physics, October 15, 2026

The chart mixes kinds of data, which is the first trap. Open X-Embodiment is robot trajectories with actions recorded. UMI data comes from handheld grippers. The 1 million hours claimed by Dyna Robotics and by Nature Will is human egocentric video, which shows a task without the robot's joint commands. An hour of video and an hour of robot action are different goods.

The denominator in the text comparison

Potter's text baseline: GPT-3 trained on about 300 billion tokens, Epoch AI (2024) estimates roughly 300 trillion tokens of quality text exist, and the largest models train on tens of trillions. He cites Chris Paxton's estimate of about 3,600 tokens per hour of video, which makes 1 million hours about 3.6 billion tokens, around 5,000 times smaller than a recent language-model corpus. Potter's headline says robot datasets are roughly 1,000 times smaller; the two figures differ by a factor of five depending on which robot dataset and which model are compared, and the essay does not reconcile them.

The rate at which data is arriving makes the gap look worse. Potter cites one source's estimate that the whole industry produces about 500,000 hours of training data per month. At that rate, 1 billion hours takes 2,000 months, or about 167 years. That is arithmetic on a rough estimate, not a forecast, but it shows why collection cost is the variable to watch.

What an hour costs

Source	Price per hour
Eidon AI (now defunct)	$60 to $100
Pantheon, UMI data	$60 to $150
Pantheon, internal pipeline	$10
Shift, pay to workers recording video	up to $20

At $10 an hour, 1 billion hours is $10 billion; at $100 an hour it is $100 billion, which matches Potter's range. Pantheon's own two figures span $10 to $150 an hour, a factor of 15, so the choice of capture method moves the bill more than a doubling of the hours would.

Why simulation does not close it

Potter points to simulation as the usual answer and gives two data points. Waymo had driven 10 million real miles and 10 billion simulated miles by 2019. OpenAI's Rubik's Cube hand trained on about 13,000 years of simulated experience yet succeeded only 20% of the time on the hardest starting configurations, and OpenAI concluded that "domain randomization alone is not enough." Paxton, quoted by Potter, adds that "neither web data nor robot data is iid, but robot data is far less iid than web data". Potter's own point is the same: "Real-world data also tends to be extremely repetitive", so an hour of robot data may carry less new information than an hour of text.

Factories as data farms

One way to avoid buying hours is to own a factory. TechCrunch reported on October 13 that Rivian's robotics spinoff Mind Robotics, which has more than $1 billion in backing, uses Rivian's plant as a high-volume environment to train and deploy its models, with Rivian as its first customer. Mind has shown no product publicly, TechCrunch says, and neither source gives Mind's cost per hour of data or its hours collected.

Other hardware stories on this site have the same missing input. Industrial Next's claim of minutes-long retooling names BMW only as an evaluator, and Boston Dynamics' 100,000-a-year Atlas target arrived with price and autonomy blank. In both cases the unit count is public and the data behind the robot's skills is not.

What would change this read

Three things. A published hour count for a major lab's training set that is far above 10,000 would narrow the gap. A measured cost per usable hour below $10 would shrink the dollar range. And a benchmark showing that video-only hours transfer to robot control at close to the rate of action-labelled hours would make the 1 million-hour claims comparable to the 4,000-hour dataset. None of the sources reviewed has published that last comparison.