The history of artificial intelligence is a history of data. Every major breakthrough has been enabled by a new source of data at unprecedented scale.
ImageNet provided the million-image dataset that fueled the deep learning revolution in computer vision. Common Crawl and The Pile provided the web-scale text corpora that enabled large language models. YouTube and other video sources provided the training data for video understanding.
Each of these data sources existed before the breakthroughs they enabled. The data was there, waiting to be used. The algorithms just needed to catch up.
Embodied AI Has No Such Data
For embodied AI, the situation is fundamentally different. The data that robots need—physical interactions with objects in real environments—does not exist anywhere. It has never been collected at scale. There is no historical archive of robotic grasping. No repository of tactile experiences. No library of manipulation trajectories.
This data must be created from scratch.
The Scale Challenge
Consider what it would take to match the scale of ImageNet in the embodied domain. ImageNet contains 14 million images, each capturing a static view of an object. Collecting these images required years of human effort, but the underlying data—photographs of objects—was already abundant. The contribution was organization, not generation.
For embodied data, each sample requires a physical interaction. A robot must actually touch an object to generate tactile feedback. A grasp must actually be performed to record its dynamics. The physical world operates in real time. There is no way to accelerate this.
A single data collection cell, operating continuously, can generate perhaps 10,000 high-quality interactions per day. At this rate, reaching 14 million samples would take nearly four years of uninterrupted operation—for a single cell. Scaling to multiple cells reduces the calendar time but multiplies the hardware investment.
And 14 million samples, while impressive for images, may be insufficient for embodied learning. Physical interactions are higher-dimensional than images, with more variables and more possible outcomes. The data requirements could be orders of magnitude larger.
The Diversity Challenge
Scale alone is not enough. The data must also be diverse—covering the long tail of objects, materials, scenarios, and tasks that robots will encounter in the real world.
Consider the variety of objects humans interact with daily. Tens of thousands of distinct object types, each with variations in size, shape, material, and state. Each combination represents a different physical interaction, with different dynamics, different tactile properties, different failure modes.
Now consider the variety of tasks. Grasping, pushing, pulling, twisting, sliding, rolling, cutting, folding, tying, inserting. Each task involves different contact patterns, different force profiles, different success criteria.
The combinatorial space is vast. Covering it meaningfully will require data collection at a scale that dwarfs anything attempted before.
The Temporal Challenge
Physical interactions are not static; they unfold over time. A single grasp produces a temporal sequence of tactile, visual, and motion data spanning hundreds of milliseconds. A manipulation sequence—folding a towel, assembling a product—may last seconds or minutes.
This temporal dimension multiplies data requirements. A dataset of 14 million static images becomes, in the embodied domain, a dataset of 14 million interaction sequences, each containing thousands of time steps. The total data volume grows by factors of thousands.
The Generational Perspective
These challenges suggest that embodied AI will not be solved in a single breakthrough. It will be a generational effort, spanning decades, with data accumulating incrementally over time.
Each year, more data will be collected. Each year, the diversity will increase. Each year, models will improve—not because of algorithmic breakthroughs alone, but because they have seen more of the physical world.
This is the long tail of physical intelligence. The common cases—simple objects, simple tasks—will be solved relatively quickly. But the tail extends indefinitely. There will always be new objects, new materials, new scenarios that robots have not encountered. Each one will require new data.
VISME’s Generational Commitment
This long view informs VISME’s strategy. We are not building for a quick breakthrough. We are building infrastructure for the decades ahead.
Our data factories are designed for continuous operation, accumulating interactions year after year. Our sensor systems are designed for durability, surviving millions of touches without degradation. Our annotation workflows are designed for scalability, processing ever-larger data streams efficiently.
We are playing the long game because embodied AI is a long game. There are no shortcuts to physical intelligence. The only path is through real interactions, accumulated over time, at scale.
The Legacy
Fifty years from now, when robots are as ubiquitous as smartphones, when they handle objects with human-like dexterity, when they fold laundry and cook meals and assemble products and care for the elderly—when all this has come to pass, someone will look back and ask: where did it start?
It started with data. With countless interactions, each one recorded and preserved. With the recognition that physical intelligence cannot be programmed—it must be learned. And with the patience to collect, over decades, the experience that learning requires.
VISME is building that legacy today. One interaction at a time.
Leave a Reply