Free Robot Data Is Not the Data You Need

In June, XDOF released ABC-130K: 3,553 hours of bimanual manipulation, 134,806 episodes, 195 tasks, free under Apache 2.0. It is the largest open dataset on the I2RT YAM arm, which is the arm a growing share of manipulation teams have standardised on. Two months earlier the Allen Institute open-sourced MolmoAct2 along with the 720 hours it was trained on, same arm, weights included.

So the obvious question. If thousands of hours of high-quality data on your exact hardware are free, why is anyone still paying to collect their own?

They are. A team I looked at this week is eleven people, six months old, publishing their own YAM teleoperation episodes one batch at a time while also fine-tuning MolmoAct2. They have the free data. They are collecting anyway.

That is not irrational, and the reason is the most useful thing I have learned about robot data collection this year.

One scoping note before going further, because the answer is different depending on who is asking. If you are one of the handful of labs training a robot foundation model from scratch, free data at this scale is genuinely valuable and you should take all of it. Everything below is aimed at the other group: the far larger number of teams building on top of someone else's base model, fine-tuning and post-training it for a specific robot doing a specific job. That is where the free data stops being enough, and that is most of the market.

What is actually free now

The floor moved this summer, and it moved a long way. A year ago, "we have bimanual manipulation data" was a defensible position for a startup. Today it is a download.

Add the open weights trained on top of them and you can go from nothing to a working manipulation policy on a common arm without collecting a single episode or signing a single contract. That is genuinely new, and anyone selling robot data who has not internalised it is selling last year's product.

Same arm is not the same setup

Here is where the free data stops.

An episode is not just an arm trajectory. It is that trajectory conditioned on a camera at a particular height and angle, a specific gripper, a table at a specific distance, particular lighting, and the specific objects in front of the robot. Change any of those and the policy that was trained on it degrades, quietly and unevenly.

You can watch this happen in public. On the ABC-130K discussion page, one researcher asks which sessions use the FlexPoint gripper rather than the standard one. Another asks for the extrinsic calibration for the wrist cameras. These are not complaints. They are the sound of someone trying to actually use 3,553 hours and discovering that you cannot condition on what is not labelled.

That is the difference between a dataset you can download and a dataset you can deploy on. The hours are real. The metadata that makes them yours often is not.

What the numbers say

The strongest public evidence I have seen on this came out of an evaluation study by Parametric, a team that runs YAM arms with their own GELLO leader-arm rig. They compared training on data collected on the same embodiment against data from a similar but different one, on a hard deformable task.

Eleven times. Not eleven percent, eleven times. On a task where contact and deformation matter, data from a nearly identical arm was worth almost nothing.

Treat that number carefully. It is one study, one task class, and deformables are the hardest case. But the direction has been corroborated everywhere I have looked this year, and it inverts the pitch the whole data industry has been making. "Our hours will transfer to your robot" is a claim that now has a published counter-example.

So what do you actually collect

It is tempting to reach for the language model analogy here, so let me be careful with it, because the loose version of it is wrong.

Pretraining did not become a commodity in language. OpenAI, Anthropic and Google all still pretrain their own frontier models, at enormous cost, and that pretraining run is arguably their central asset. What changed is that almost nobody else has to. For the thousands of companies building on top, the base model became something you rent or download, and the money moved to the layer above it: instruction tuning, preference data, evaluation, retrieval over their own corpus.

So the honest framing is not that pretraining is commoditised. It is that pretraining is trending toward commoditisation for everyone who is not building the base model themselves.

Robotics is early in the same split, and the split matters more than the trend.

If you are building a robot foundation model, general data is still your core asset. You need vastly more of it than exists publicly today, ABC-130K and MolmoAct2 are a useful floor rather than a substitute, and every additional open corpus genuinely helps you. There are maybe a dozen organisations worldwide in this position.

If you are deploying a policy on top of someone else's base model, the calculus inverts. The base is increasingly something you download. Free hours, open weights, someone else's capital. Competing there is competing with a Hugging Face URL.

The asymmetry between those two groups is the part people skip. Foundation model builders are a handful of well-funded labs. Teams fine-tuning and post-training an existing model for a specific robot, a specific cell, a specific customer, number in the hundreds and are growing much faster. Almost everyone reading this is in the second group, and almost every dollar of robot data spend will eventually come from it.

For that second group, and only that group, the expensive part is what comes after: a few hundred hours of your task, in your scene, with your gripper and your camera rig. Small in volume, and where the reliability actually comes from.

This is why an eleven-person team with free data on their shelf still runs teleoperation sessions. They are not replacing ABC-130K. They are doing the thing ABC-130K cannot do for them.

The budgeting consequence is the same either way, and it is uncomfortable for anyone selling volume: the expensive data is not the biggest data. It is the most specific.

What this means if you are buying

Three questions worth asking any robot data vendor:

1. What is in this that I cannot download? If the answer is hours on a common arm doing common tasks, the answer is nothing.
2. What is the metadata? Gripper variant, camera extrinsics, lighting, object set, per-episode. Unlabelled hours are hours you cannot condition on, which makes them pretraining filler.
3. Does it match my embodiment, or does it merely resemble it? 77 against 7 is the price of getting that wrong.

The free data is real and you should take it. If you are training a base model it may be the most valuable thing you download this year. If you are building on top of one, it improves performance up to a point. What it will not do is carry you to a deployment-ready standard, and the gap between those two things is wider than the download makes it look.

Related: The Data Wall on why demonstration data is the bottleneck at all, and How Robots Learn From Us on where these demonstrations come from in the first place.