back to blog

Robotics Needs a Better Data Diet

Robotics Needs a Better Data Diet

A lot of AI progress still gets narrated as a scale story.

More data. Bigger model. Broader mixture. Hope the mess averages out into competence.

Sometimes that works. But one of the more interesting new papers on arXiv, Psi-Zero: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation, argues that for humanoid robots this instinct can point in the wrong direction.

The paper's claim is not that scale stopped mattering. It is that embodiment matters enough to break naive scaling recipes.

The authors argue that co-training on a giant mixed soup of human and humanoid data is suboptimal because humans and humanoids do not move the same way. Their kinematics differ. Their constraints differ. Their action spaces differ. If you blur those differences together too early, you get an expensive compromise: lots of data, weak transfer, and control that still feels brittle.

So they split the problem in two.

First, they pre-train a vision-language-action backbone autoregressively on large-scale egocentric human videos. That stage is about learning broad visual-action structure: what hands are doing, how objects move, what manipulation looks like from a first-person view, how tasks unfold over time. Then they post-train a flow-based action expert on high-quality humanoid robot data to learn the thing that actually makes or breaks robotics: precise control in a real body.

That sounds obvious after the fact, which is usually a good sign.

The better result in the paper is not just that this works. It is that the data recipe is cleaner than the usual scaling fantasy. Psi-Zero reports its best results with about 800 hours of human video and 30 hours of real-world robot data, beating baselines that were pre-trained on more than 10 times as much data by over 40% in overall success rate across multiple tasks.

That is a useful result because it points to where the bottleneck might really be.

Maybe the path to more capable robotics is not "scrape everything that moves." Maybe it is learning which parts of the problem are universal and which parts are body-specific, then training those layers differently.

This matters beyond humanoids.

A lot of the next wave of AI systems will have the same shape. They will need a broad prior from internet-scale or human-scale observation, then a narrower adaptation step that binds that prior to a specific interface with reality: a robot body, a software environment, a tool stack, a lab workflow, a vehicle, a factory cell. General intelligence in practice may look less like one monolith and more like a good base model attached to a well-chosen adapter for the world it has to survive in.

That is why this paper feels more important than yet another benchmark win.

It is really a paper about respecting contact surfaces.

In language models, people are slowly learning that verifiers, sandboxes, and environments matter because they give the model something firmer than its own priors. In robotics, the equivalent lesson is that the body is not a rounding error. You do not get robust physical competence by pretending a humanoid is just a human video model with motors taped on.

You need a bridge between generic understanding and embodied execution.

Psi-Zero's training recipe is a proposal for building that bridge on purpose.

There is also a bigger accelerationist point hiding in here.

If you want a more abundant future, you do not just want smarter models in the abstract. You want intelligence that can be instantiated cheaply across many useful surfaces. Screens, tools, vehicles, labs, homes, warehouses, machines. That means the winning systems will probably be the ones that can inherit a lot of general structure, then specialize fast without needing absurd amounts of bespoke data every time.

A recipe like this pushes in that direction. Human video supplies cheap, broad priors. Small amounts of high-quality robot data do the last-mile binding. If that pattern keeps working, it lowers the cost of turning general perception into real capability.

That is how the future gets less artisanal.

The paper also says it will open-source the whole stack: the data processing and training pipeline, the humanoid foundation model, and a real-time action inference engine. If that actually ships in usable form, it matters almost as much as the headline result. Open robotics infrastructure has a compounding effect. It turns isolated demos into surfaces other people can build on.

And that is the real taste test now.

The field has enough glossy robot clips. What it needs is transfer. Reproducibility. Recipes that tell you which data to collect, which distinctions to keep, and where brute-force scaling is just hiding confusion.

Psi-Zero is interesting because it makes a narrower and probably truer claim than the usual frontier theater. It does not promise universal robotics because someone trained on a mountain of video. It says something more disciplined: if you separate broad visual-action learning from embodiment-specific control, you can get more capability out of less data.

That is a much better kind of progress.

Not because it is modest, but because it is legible.

And legible progress is what scales. Once you know which parts of intelligence should be shared and which parts should be locally adapted, you can start building systems that spread much faster through the physical world.

That is when robotics stops looking like a collection of expensive prototypes and starts looking like infrastructure.