Scaling Laws for Robotics: Does More Data Win?

Pool over a million robot trajectories from twenty-two embodiments, train one policy on the whole pile, and the striking part is not that it works. It is how modestly the numbers move. That was a lesson of Open X-Embodiment, the 2023 collaboration that merged more than 60 datasets from 21 institutions into a single training mixture. Co-training produced real positive transfer, but nothing like the clean, predictable climb a language model gives you every time you feed it more text.

The question underneath is easy to state and hard to answer. Large language models obey scaling laws: supply more parameters, more tokens, and more compute in the right ratio, and loss falls along a power law you can extrapolate for years. Does the same hold for robot policies? If Physical Intelligence, a dozen humanoid labs, and Roborecs are all betting that more demonstration data buys better behavior, it is worth asking precisely where that bet pays and where it quietly stalls.

The honest answer: more data helps, but the robotics curve is governed by something the language curve mostly ignored, which is coverage. A robot policy is only as good as the slice of the world its data touched, and the world of physical contact is far wider, and far more expensive to sample, than the text sitting on the internet.

What the language scaling laws actually promised

The 2020 Kaplan results and the 2022 Chinchilla correction handed the field a formula. Test loss dropped as a power law in model size, dataset size, and compute, with the three needing to grow together. Why it worked is worth stating plainly. Language pretraining is self-supervised, so every sentence on the web is free supervision. The objective, predict the next token, is identical at train time and test time. And the corpus already existed at a scale of trillions of tokens; nobody had to manufacture it.

Robots inherit none of those three properties for free. There is no trillion-trajectory corpus of manipulation sitting on a server. The training objective, imitate an action, is not what the robot faces at deployment, where it has to recover from its own mistakes. And every unit of data costs real hardware time.

Where the analogy does hold

It is not all divergence. The most productive idea of the last three years is that a robot policy can borrow a language or vision backbone that already paid for its own scaling. Vision-language-action models, including RT-2, Gemini Robotics, Physical Intelligence's π0, and NVIDIA Isaac GR00T, start from a web-pretrained VLM and graft an action decoder on top. The semantic prior, what a mug is, what "wipe the spill" means, arrives for free, and RT-2 showed that some of this semantic generalization carries into the robot's behavior. The policy only has to learn the last mile of motor grounding.

Cross-embodiment pooling transfers too, up to a point. RT-X showed a policy trained across many robots beating single-robot baselines. DROID, roughly 76,000 demonstrations across 564 scenes gathered at 13 institutions, was built so that scene and task diversity, not raw volume, is what feeds the model.

Where it breaks

Now the joints where the analogy cracks.

Errors compound. A language model that mispredicts one token moves on. A policy runs closed-loop: one small error shifts the robot into a state its training data never covered, the next action is worse, and the trajectory drifts off the data manifold. This covariate shift means real-world performance can degrade faster than any offline loss curve would predict.

The data is not fungible. One more token is interchangeable. One more trajectory from a different gripper, camera placement, or control frequency is not. The embodiment gap means data from a Franka arm does not cleanly transfer to a humanoid hand, so you cannot simply add everything together and expect the pile to help.

Someone pays for every second. Teleoperated demonstration runs in real time on real hardware. There is no scraping shortcut, and no synthetic corpus that fully substitutes for contact.

What actually moves the needle

If volume alone is not the lever, what is? The evidence from the last two years points at diversity and distribution rather than count.

In robotics the useful axis is rarely how many demonstrations you have. It is how much of the deployment world they touched, and how little of it you left unseen.

Three levers keep showing up. Diversity of scene and object, so the policy has met contact under many conditions. Quality and consistency of demonstration, since a policy imitates its teacher's habits, the good ones and the bad. And targeted coverage of the long tail, the rare states where deployed robots actually fail. Toyota Research Institute's Large Behavior Models work argues that broad multitask pretraining makes each new skill cheaper to add, a compounding return that piling up single-task data does not give you.

The human-data shortcut

This is where human egocentric video enters. A person cleaning a kitchen produces a dense stream of first-person contact, dexterity, and intent, at a cost far below teleoperating a robot through the same motions. Ego4D assembled thousands of hours of such footage; Ego-Exo4D added time-synchronized first- and third-person views of skilled tasks. The bet across the field is that this data covers the human long tail cheaply, and that the gap between a human hand and a robot hand can be closed with the right retargeting plus a smaller pile of on-robot data to anchor it.

It is not a free lunch. Human video lacks robot proprioception and clean action labels, and the morphology gap is real. But as a way to buy coverage, the axis that actually bounds robot performance, it is far cheaper per hour than any teleoperation rig.

So, does more data mean better policies? Yes, but the word doing the work is not "more," it is "which." Language models could be lazy about distribution because the web already covered most of what they would ever be asked. Robots have no such luck. The teams that win the next few years will not be the ones with the largest trajectory count. They will be the ones whose data most closely matches the messy, contact-rich, long-tailed world their robots are sent into.