Part 8, the capstone of our series on the data problem in Vision-Language-Action research. [Part 1: the bottleneck overview · Part 2: the language gap · Part 3: the failure data · Part 4: the evaluation crisis · Part 5: the synthetic trust layer · Part 6: labeling human video · Part 7: the greenfield modalities]
Robotics inherited a belief from language modeling: scale wins. Add more data, add more parameters, and capability follows. It's a good prior, it built the models that made this series worth writing, and for a while it drove VLA progress too. But the most careful recent work in the 1,228-paper corpus we analyzed has started to find its edge. On the margin that matters, more robot data is no longer the reliable lever the field assumed, and the reason ties every earlier part of this series together.
This is the argument that unifies the other seven posts. Language, failure, evaluation, verification, video labeling, and multimodality all look like separate problems. They are the same problem, and the scaling results are what prove it.
It's worth being fair to the scale-first view, because the field had every reason to hold it. Robotics is growing at a rate that makes "just add data" feel obviously correct. In the broader scan of 53,800 robotics papers this corpus was drawn from, the field grew 127% in a single year. Vision-Language-Action, the fastest-moving corner of it, grew 5× year over year, more than three times the pace of the next-fastest area. When a field is compounding that fast, the reflex is to feed it: collect more trajectories, teleoperate more hours, pool more datasets, and ride the curve the way large language models rode the web.
The reflex worked for a while, for the same reason it worked in language: early on, almost any additional data raises the pretraining floor. The question this final post takes up is what happens after that phase ends, when you have a specific robot, a specific task, and a specific deployment, and you are deciding what data to collect next. That is where the corpus says the scaling story changes.
Start with the evidence that naive scaling has stopped paying off, because it's more specific than the usual hand-waving about diminishing returns.
Negative transfer. A 2026 study on rethinking VLA scaling (arXiv:2602.09722) found that naively pooling heterogeneous robot datasets often induces negative transfer: adding data from other robots can make your model measurably worse. This is the sharpest possible rebuke to volume-thinking: more data isn't neutral-but-wasteful, it's actively harmful when it's the wrong data. The sign on the marginal episode can be negative. A team that assumes every additional dataset helps is running a real risk of paying to degrade its own model.
The 5% coreset. A data-distillation result (arXiv:2511.16233) showed that a carefully selected 5% coreset recovers 85–90% of full-dataset performance. Read that the other way around: 95% of the volume in a typical corpus is doing perhaps 10–15% of the work. The value isn't spread evenly across the episodes; it's concentrated in a small, well-chosen subset, and the rest is ballast. If you can identify the 5% that matters, you can match a competitor's model at a fraction of the collection and training cost. Identifying it is a curation and annotation problem, not a collection one.
Persistent long-horizon failure. MPVI (arXiv:2606.00985) documents long-horizon manipulation failures that persist despite fine-tuning on large teleoperated datasets, concluding that more data alone may not resolve them. Some capability gaps are structural: they're about what kind of data you have, not how much, and no amount of the same thing closes them. You can quadruple a dataset that lacks corrective and failure signal and still have a policy that cannot recover from its own mistakes, exactly the gap Part 3 identified.
Put together, these say the field has exhausted the easy gains from volume. Scale still helps for broad pretraining priors. But for a specific deployed capability, the lever moved.
If volume isn't the lever, what is? The corpus gives a strikingly consistent answer: diversity, deliberately constructed.
The cleanest result is SEVO (arXiv:2605.11114), which found that a diversified collection protocol (systematically varying lighting, backgrounds, and distractors during teleoperation) was the single most important factor for generalization, while in-distribution-only data produced near-zero transfer to new environments. Not more episodes. More varied episodes. A smaller, deliberately diverse dataset generalized where a larger, homogeneous one did not. The imitation-learning literature says the same thing from the collection side: what these methods consume is clean demonstrations broken into the right sub-steps, with quality and preference labels attached, not raw hours.
This is the through-line of the entire series, stated as a scaling law. Every earlier part is a special case of "diversity and quality beat volume":
Part 2's language finding, relabeling trajectories for diverse instructions moving success from 0% to 90%, is diversity applied to the language modality. Part 3's failure data is diversity of outcome: a corpus that contains only success is homogeneous in the dimension that determines robustness. Part 4's benchmark collapse is what homogeneity looks like at evaluation time: a model that memorized a narrow distribution and can't leave it. Part 5's trust layer exists because synthetic volume without verification is volume of unknown quality. Part 6's video-labeling problem is that a million unlabeled hours is enormous volume with near-zero usable signal until annotation adds the diversity of structure. Part 7's greenfield modalities are diversity of sensor: the coverage the corpus is missing entirely.
Seven posts, one finding: the axis of progress turned from how much to how good and how varied.
If composition is the lever, it helps to be concrete about what composing a dataset involves, because it is a different discipline from collecting one. Collection asks "how many more hours can we record?" Composition asks a sequence of harder questions. Which slices of the data are actually carrying the model's performance, and which are ballast you could drop? Where is the distribution thin, in lighting, in object variety, in failure modes, in phrasings, in sensor coverage, and what targeted collection would fill those specific holes? Which episodes contradict each other or come from robots different enough to cause negative transfer? Which trajectories are worth the expense of fine-grained language, failure, or tactile annotation, and which are fine left coarse?
None of those questions is answered by a bigger collection budget. Each is answered by measurement, curation, and human judgment applied to the data you already have. That is the work the scaling results are quietly pointing at, and it is the work the field's own surveys now call for by name.
It would be easy to read "scaling is hitting limits" as bad news. It's the opposite, and it's worth being clear about why.
If capability were purely a function of volume, robotics data would become a commodity race, won by whoever can afford the most robots and the most operator-hours, a game of capital, not craft. The scaling results say that's not the game. A carefully composed 5% of the data can match 90% of the performance; a diversified collection protocol beats a larger homogeneous one; the wrong extra data can hurt. That means the decisive skill is composition, choosing, diversifying, annotating, verifying, and covering the hard cases, not accumulation.
That's a more interesting field to work in, and a more open one. It rewards judgment over budget. It means a team with less data but better data can win. And it means the surveys quoted in Part 1 are right in a deeper way than they may have intended: when they call for treating data infrastructure as a first-class research problem, the scaling results are the proof that the infrastructure, the pipeline that selects, diversifies, annotates, and verifies, is where the remaining performance actually lives.
Across 1,228 papers, the field diagnosed its own bottleneck, discarded the obvious fix, and converged on a subtler one. The bottleneck is data. The obvious fix, more of it, has stopped working. The real fix is data that is diverse, richly annotated, failure-aware, honestly evaluated, multimodal, and verified: every adjective an annotation and curation problem, and every one of them a quality problem rather than a quantity one.
The field spent its first VLA era asking how to collect more. Its next era is about learning to collect better. The evidence for that shift is already in the literature, hundreds of papers deep, inside a field growing 127% a year that can no longer buy its way forward on volume alone. The teams that internalize it first won't be the ones with the most data. They'll be the ones who figured out which data was worth having.
This concludes our eight-part series on the data problem in Vision-Language-Action research. It draws throughout on our analysis of 1,228 VLA papers from arXiv cs.RO (Feb 2023 – Jun 2026), part of a broader scan of 53,800 robotics papers. See Part 1 for full methodology.