Robotics May Have Found Its Scaling Law — And It’s Not About Robot Data

To understand why this matters, consider what happened in language AI. The central training objective behind many large language models is almost absurdly simple: Given everything so far, predict what comes next. A word. A subword. A token. Scale that objective across enormous datasets and increasingly large models, and surprisingly general capabilities emerge.

The important part was not only the algorithm. It was the existence of a nearly inexhaustible supply of training material. The internet was already there. Robotics does not have an internet of actions. There are millions of videos of people making beds. But most conventional robot learning systems cannot simply watch those videos and immediately know which motor commands should be sent to a humanoid robot. A video tells you what happened visually. A robot needs to know what to DO. That difference created one of embodied AI's biggest bottlenecks.

Google's Open X-Embodiment effort, for example, brought together data from more than 20 research institutions, covering 22 robot embodiments and more than one million episodes. The project itself illustrates both the promise and difficulty of aggregating robot experience: no single laboratory can easily generate enough diverse physical data alone.  Even a million robot episodes is tiny compared with the data regimes of frontier language models. And unlike a text token, every new robot trajectory can be expensive. That is why a question increasingly hangs over the robotics industry: What if robots don't need to generate most of their own training data?

Figure's answer is a dataset called Index. Instead of requiring a humanoid robot to collect every demonstration, Figure built a network for gathering videos of humans performing physical tasks in the real world. As of August, the company said its Index app had been downloaded more than 264,000 times across over 100 countries, with more than 44,000 weekly active contributors. Those users had uploaded more than 16 million videos.

By September, Figure said the system was receiving roughly 35 minutes of new human experience every second. These are company-reported figures, but the scale illustrates what the company is trying to build: a data collection system whose growth is no longer constrained by the number of robots Figure owns. That changes the economics dramatically. A robot fleet can grow linearly. A consumer data network potentially grows much faster.

The idea is not that a human hand and a humanoid robot hand are identical. They obviously are not. Human bodies have different joints, proportions, dynamics, tactile feedback, and control systems. The bet is that underneath those differences lies a large amount of transferable structure. A human reaching into a cupboard teaches more than a particular arm trajectory. The demonstration contains information about:

• which objects can be grasped;

• how objects interact;

• what parts of a scene matter;

• how bodies navigate around obstacles;

• how a long task decomposes into smaller actions;

• and how the physical world tends to respond when something moves.

In other words, much of the useful information may exist at a level above the exact motor command.

Research outside Figure is increasingly exploring the same idea. A 2026 survey of robot learning from human video identifies scalable human video as a potential way around the robot-data bottleneck, while noting that transferring skills across the human-robot “embodiment gap” remains a major open problem. NVIDIA researchers have also demonstrated positive transfer from human manipulation videos through latent-action pretraining, while NVIDIA's GR00T stack now explicitly combines real robot data, synthetic data, and internet-scale video. The emerging recipe starts to resemble foundation-model training: Learn broadly first. Specialize later.

Figure tested Helix 2.5 on three household behaviors: tidying a living room, folding towels, and making a bed.

The company then took the same trained policies into 30 previously unseen homes in the San Francisco Bay Area. According to Figure, no training data was collected inside those evaluation homes, and the robot had not previously seen the particular toys, towels, or bedding used in the tests. This matters because many impressive robotics demonstrations operate under much narrower conditions. A policy may work extremely well on a particular table, in a particular laboratory, using objects similar to those encountered during training.

• Change the lighting.

• Move the table.

• Replace the object.

• Change the room.

• Performance can collapse.

Helix 2.5 was designed to test whether broad pretraining could reduce that brittleness.

Figure conducted an ablation that makes the result especially interesting. Two policies received the same task-specific training data and used the same architecture, optimization setup, and evaluation. One started from random weights. The other started from the Index-pretrained model. The policy trained from scratch achieved a 9% zero-shot task success rate. The Index-pretrained version reached 56%. Figure required the entire task to be completed for a success; partial completion did not count.  That is a substantial difference but the more consequential experiment came next.

Figure trained four models using progressively larger subsets of Index, spanning an 8× increase in human pretraining data. Model size remained fixed. Downstream task training remained fixed. Only the amount of human pretraining data changed.

Figure reports that held-out robot-action prediction loss declined smoothly as the Index dataset grew. Using the smaller training runs, the team says it could forecast the loss of the largest run to four decimal places, with forecasting error equivalent to just 0.54% of the variation measured across the full 8× range. This is where the robotics story starts sounding strangely familiar.

Years ago, researchers studying language models discovered that model performance often improves according to surprisingly regular relationships between data, compute, and model size. Those scaling laws changed how AI systems were built. Instead of asking: “Will a much larger model work?” researchers could increasingly ask: “Given this much compute and data, how much better should it become?” That predictability helped turn frontier AI training from an experimental art into something closer to an engineering discipline. Figure is suggesting that part of the same phenomenon may exist in robotics. 

Not: more robot demonstrations → better robot. 

But potentially: more human physical experience → better robot action prediction.

That distinction is enormous.

The rise of LLMs made the token one of the most economically important units in computing.

It is worth asking what its equivalent will be in embodied AI. For years, the obvious answer seemed to be the robot trajectory:

observation → action → next observation

A robot sees the world, performs an action, observes what happened, and learns from the transition. The problem is that robot trajectories are expensive. Human activity is not.

Billions of people are continuously opening doors, cooking food, moving boxes, using tools, folding clothes, assembling products, cleaning rooms, and interacting with almost every physical environment imaginable. Those actions represent an enormous reservoir of physical intelligence. The challenge is encoding them in a form robots can use. If that problem can be solved, then the unit of robotic scaling may turn out to be something much broader than a robot trajectory. It could be a human behavior segment. Or a learned latent action. Or some compressed representation of how bodies, objects, and environments change over time. The exact representation remains unsettled.

But the strategic implication is clear: The winning robotics company may not be the company that owns the largest robot fleet. It may be the company that builds the largest useful model of physical experience.

There is an important qualification. Figure has not discovered the first evidence that robot learning improves predictably with data. 

Earlier academic work has already found scaling behavior in robotic manipulation. A 2024 study examining more than 40,000 demonstrations and 15,000 real-world robot rollouts found that generalization performance exhibited an approximate power-law relationship with the diversity of training environments and objects. One of its most interesting findings was that diversity mattered more than simply collecting more repetitions. After enough demonstrations had been gathered for a given environment or object, additional examples produced diminishing benefits. Expanding the number of environments and object variations was more valuable.

That finding is important because it suggests that “more data” is an incomplete description of robotic scaling. The real variable may be: more diverse physical experience. Figure's more specific claim is that it has observed a scaling law for human-to-humanoid transfer on a humanoid robot. That is interesting precisely because it moves the expensive part of training away from the robot.

The analogy to language models is tempting. For LLMs, datasets such as Common Crawl helped turn the public web into raw material for machine intelligence. Physical AI may need something comparable.

But unlike text, physical experience is not naturally represented as a clean sequence of discrete symbols. Video contains huge amounts of irrelevant information. Actions may be partially hidden. Camera viewpoints change. Human hands do not map neatly onto robot end effectors. Videos often do not reveal forces, tactile feedback, object weight, or the precise intention behind a movement. So simply feeding YouTube into a robot will not solve robotics. The real breakthrough would be learning representations that extract the actionable structure hidden inside that video. That is exactly why human-video learning, latent actions, video world models, and cross-embodiment learning are converging into one of the most important research areas in robotics.

NVIDIA, for example, has publicly argued that conventional teleoperation fundamentally limits data scale because a robot must remain in the collection loop. Its researchers are exploring human egocentric video and video models as ways to train robots without requiring physical robots for every new piece of experience. Its DreamGen research pushes the idea even further: use video world models to generate synthetic robot trajectories, effectively attempting to substitute additional GPU computation for some of the physical data collection traditionally required from humans and robots.

Taken together, three potential scaling engines are emerging: Human experience, synthetic experience and robot experience. The future foundation model for robotics will probably use all three. The unanswered question is their optimal ratio.

There is also a reason to be careful with the word “scaling law.” Figure's result is still limited.

It involves four training runs over an 8× data range. The reported scaling relationship concerns held-out action-prediction loss, not a demonstrated power law connecting data directly to real-world task success. The household evaluation covers three task families and 30 homes in one geographic region. And both the scaling experiment and the zero-shot evaluation were conducted and reported by Figure itself.

Independent replication will matter. More importantly, a model that succeeds 56% of the time is fascinating research, but it is not yet a reliable household worker. Robotics faces problems language models do not. A chatbot can produce a bad sentence and try again. A robot dropping a glass, colliding with furniture, damaging equipment, or hitting a person does not get the same luxury. 

Physical intelligence ultimately has to deal with force, contact, friction, deformable objects, uncertainty, wear, latency, and safety. Scaling can improve a model's prior understanding of the world. It does not eliminate physics.

Still, something fundamental may be changing. For decades, robotics was built around teaching a machine a task. Collect demonstrations for one behavior. 

• Train it. 

• Deploy it. 

• Encounter another environment. 

• Collect more demonstrations. 

• Train again.

That workflow resembles machine learning before foundation models. The emerging alternative is almost the reverse: Pretrain on the physical world first. Then teach the robot what job you want it to perform. If that paradigm works, a future household robot may not learn to make your bed by repeatedly practicing on your bed.

It may arrive already understanding thousands of ways beds, sheets, pillows, rooms, bodies, and fabrics behave—because it has indirectly learned from millions of humans interacting with the physical world.

The final few examples would specify the task. The billions before them would provide the intelligence. This is what made foundation models transformative in language. And it may now be starting to happen in robotics. The most important resource in the next phase of embodied AI may therefore not be robots. It may be human experience itself.

Text was the fuel of the first foundation-model era but human behavior could become the fuel of the next one.