Gasgoo Munich- How should we react when a company announces it has amassed 1 million hours of robotic data?
Don't let the "1 million" figure dazzle you just yet.
During a media roundtable on August 31, Gasgoo Embodied Intelligence put this question directly to Yao Maoqing, chairman and CEO of Maniformer: Beyond the sheer hours, how do we determine if a company's embodied data holds real value?
His response threw cold water on the "million hours" metric: "Blindly collecting 1 million hours is honestly quite easy—you just repeat the same tasks. But it holds almost no value."
This is likely the real challenge the embodied data industry must confront after the hype of "million hours."
As discussed previously, if robots are to enter the open world, they require data on a scale far beyond what exists today. We also noted that as demand shifts from millions to tens or hundreds of millions of hours, data production is expanding from collection factories to communities, households, and actual workplaces.
Yet once that scale truly arrives, a new problem emerges quickly: having data is not the same as having training data.
If an employee wears a device while working for three hours, the hard drive will naturally gain three hours of video.
But within those three hours, was the person actually performing the task? Did their hands leave the frame? Was the motion trajectory accurate? Had the scene already been captured countless times? And what, exactly, can the model learn from it?
If these questions go unanswered, embodied intelligence risks swinging from "data scarcity" to the opposite extreme: possessing a sea of data but having no idea what is worth training on.
So, beyond the million-hour mark, the real competition in the data industry likely won't be about "collection," but "refining."
"100 Hours of Video" Is Not the Same as "100 Hours of Training Data"
The most common misconception in embodied data is equating camera runtime with data assets.
In reality, the two are far apart.
When explaining "effective data" during the roundtable, Yao offered several basic, even trivial, examples:
If the hands aren't in the frame during capture, it's useless. If the person wears the device but "slacks off" with no movement for long periods, that doesn't count either. Even when the person is operating, the task itself must hold meaning.
Only then do we get to the more technical requirements.
If a device claims 60 FPS, can the actual average frame rate truly reach 59.98 or 59.99? Is the camera calibration accurate? Is the hand trajectory extracted via HandPose off by several centimeters, or compressed to 5–7 millimeters? More importantly, that precision must be verifiable against ground truth.
In other words, just because something happened in reality doesn't mean it was recorded correctly.
This is a crucial distinction between embodied data and standard internet video.
Even if a cooking video drops a few frames, a viewer can still understand that the chef is chopping vegetables.
But if a robot needs to learn movements from it, a few frames of error could mean deviations in hand position, moment of contact, or even object state.
What looks "close enough" in an image may not be "close enough" inside a robot's control system.
The true production process for embodied data, therefore, only begins after the camera stops recording.
The MEgo Engine unveiled by Maniformer aims to bridge the gap between raw data and training datasets: Raw data must undergo preprocessing, spatial perception, atomic semantic annotation, and quality assessment, while also completing time alignment, trajectory reconstruction, human keypoint estimation, and task decomposition.
Regardless of how well this system performs, it illustrates at least one point: so-called embodied data companies likely cannot remain mere "video recording companies" in the future.
The real barrier is translating a segment of human behavior into training assets that a model can truly understand and deploy.
Robots Don't Learn "Wiping a Table"—They Learn the Actions Within It
What does this translation actually entail?
Imagine a cleaner wearing a device and working continuously for two hours.
To a human, this is a complete, natural workflow.
He walks to a table, picks up a rag, moves a cup, wipes the surface, tosses trash into the bin, and moves to the next table.
But the model struggles to extract specific insights from a two-hour video:
At what exact moment does the task begin?
Which segment involves grasping?
When does the placement occur?
Which action constitutes wiping?
And after which failure does a readjustment happen?
The first step, therefore, is often to "disassemble" long-duration human behavior.
Liu Li, Maniformer 's technical lead, explained that their system decomposes hours—or even 2 to 3 hours of continuous data—into hierarchical semantic information. It breaks down overall tasks into subtasks, and further into atomic actions. Currently, the system can identify over 120 types of actions, such as placing, releasing, positioning, and grasping.
This process reveals what the truly valuable unit of embodied data might actually be.
It is likely not "hours."
It could be a single grasp.
A single insertion.
A single rotation.
A retry after a failure.
Or even a complete interaction chain of "reach—contact—move—release."
This is also why simply comparing which company has 1 million hours versus 2 million hours is increasingly meaningless.
One dataset might be packed with hundreds of thousands of hours of repetitive motions; another, while smaller in total volume, could contain a vast array of objects, environments, and reusable atomic skills.
If the latter is what ultimately enters the model, then "hours" is likely just a production metric, not a value metric.
Yao also noted during the roundtable that clients currently prioritize the scenario, followed by spatial precision like HandPose, and then the accuracy of action and label descriptions. The ultimate goal is for data to feed directly into the model while ensuring the motion performance mirrors normal human operation.
This means the embodied data industry is undergoing a typical shift: from "do we have it?" to "can we use it immediately?"
Even "Bad Data" Might Not Be Truly Bad
Yet simply deleting all low-quality data doesn't solve the problem either.
This point is actually one of the more compelling technical concepts from Maniformer's release.
Within its ManiEval data quality control system, Maniformer proposes a principle not of simple "filtering," but of: "no filtering, only grading."
According to the presentation, data is scored across dimensions such as data type, task type, sensor quality, semantic quality, and action quality. It then flows into four distinct uses: precise data for policy imitation learning; data rich in semantic and environmental information for pre-training and representation; data containing environmental interference, motion deviations, or partial failures for error detection and task progress perception; and long-tail boundary data for generalization testing and enhancing corner case capabilities.
This raises a question worth exploring: What constitutes "bad data"?
If a person's hand slips while grabbing a cup, from the perspective of imitation learning, that is certainly not a standard motion worth replicating.
But once robots enter the real world, will they never slip?
If a model has never seen failure, how will it recognize when failure is occurring?
Furthermore, if a robot's grip shifts halfway through moving a cup, should it continue, re-grasp, or abort the task entirely?
In such moments, a segment of "unsuccessful" human motion may be far more valuable than 1,000 identical successful attempts.
This logic mirrors the evolution seen in the autonomous driving industry.
In the early days, everyone was obsessed with comparing test mileage.
But as system capabilities improved, what became truly scarce was no longer 10,000 kilometers of normal straight-line driving, but those rare, difficult, or system-failing corner cases.
Embodied intelligence will likely undergo the same transition.
After a robot folds the exact same shirt 1,000 times, the 1,001st success adds little new knowledge.
But if it suddenly encounters a shirt that is exceptionally soft, slippery, or irregularly shaped—and fails—that is different.
That failure might actually be the data point worth keeping.
Consequently, the future metric of value for the data industry may shift from "how many successful actions were collected" to: how many previously unknown problems were discovered.
The Most Valuable Data Might Be the Moment the Robot Fails
Pushing this logic further leads us to the crux of the matter:
Data is not a one-time delivery to a model.
A truly mature data ecosystem should form a closed loop with robot deployment.
Maniformer demonstrated a corresponding approach: a model gains initial capabilities from early data and enters real-world execution. The data generated by the robot flows back, the system identifies failure and difficulty scenarios, reintroduces them into training, and then updates the model policy.
The REMORA Value Model attempts to further assess task progress, detect anomalous actions, and reuse failure samples as high-value data.
Of course, a degree of skepticism is warranted here.
This closed loop is currently primarily a technical roadmap disclosed by Maniformer, backed by their own test results. Whether it delivers stable gains across different robots, models, and complex real-world deployments remains to be verified by external projects.
But the direction itself is critical.
Because it implies the embodied data industry could ultimately shift from a "inventory business" to a "feedback business."
The old logic was:
Collect 1 million hours, organize it, and sell it.
The future may look like this: model trains — robot works — failure is found — cause is determined — corresponding data is recollected — model retrains.
At this stage, what a robotics company truly needs is not "how many hours are left on the supplier's shelf."
It is: Which specific data point does my robot lack most right now?
The difference between these two questions is vast.
The former organizes production based on inventory.
The latter organizes production based on model capability gaps.
If this transition holds, the core competency of the embodied data industry will likely change as well.
Being able to collect 1 million hours in one go is certainly a capability.
But the truly difficult task is this: after a robot fails in an unfamiliar scenario, can you quickly pinpoint:
Why did it fail?
Is it lacking visual understanding?
Spatial trajectory?
Experience with a specific object interaction?
Or has it simply never encountered this boundary condition?
And then return to the real world to accurately fill that experience gap.
The true "flywheel" in the data flywheel isn't just accumulating more data—it's the model increasingly knowing what it lacks.
The Data Industry Must Ultimately Answer One Question: What About the Results?
So, circling back, how do we judge if an embodied data company's data holds real value?
Yao's response during the roundtable actually had three layers.
First, look at the breadth of coverage—you can't spend 1 million hours repeating the same few job types. Second, see if they can produce detailed quality verification reports. Third, check if the data has actually been accepted by clients, signed off on, and generated revenue, because market payment itself is a form of validation.
These metrics are all reasonable.
But from an industry observation standpoint, there should be one final layer: model performance.
Data is not ultimately for selling data.
Nor is it for producing a glossy quality report.
It should ultimately answer this:
After the model uses the data, what actually changes in the robot?
Did the success rate improve?
Did it learn tasks it previously couldn't handle?
Can it still function in an unfamiliar scenario?
Can it recover after an error occurs?
If adding 10,000 hours of data boosts the success rate by 10 percentage points with Dataset A, but only 1 point with Dataset B, then even if both meet 60 FPS, millimeter-level precision, and have complete labels, their true value to the model differs significantly.
This is a piece of public evidence still missing across the entire embodied data industry.
Maniformer showcased capabilities such as HandPose, data governance, ManiEval, and the Value Model, releasing a series of technical metrics. For instance, according to the company's test results, its HandPose achieved a hand reconstruction error of 4.29 millimeters and a temporal inter-frame jitter error of 1.11 millimeters during specific occlusion, interaction, and high-speed movement tests.
These numbers can demonstrate a portion of the data processing pipeline's capabilities.
But it is still not the final answer.
Does reconstructing the hand's position more accurately ultimately help the robot perform the task better?
That is the moment when data value truly closes the loop.
Today, the embodied intelligence industry is keen on discussing Scaling Laws, which is easy to understand.
Millions, tens of millions, hundreds of millions of hours—these are impactful, easily shareable figures.
But as data scales up, the industry will inevitably enter a new phase:
"How much data you have" will become increasingly less important, while "what the robot actually learned from each additional piece of data" will become increasingly vital.
This may well be the true watershed for the embodied data industry.
The first stage is about who can collect.
The second stage is about who can scale production.
But by the third stage, the competitive question may become:
Who knows sooner what the robot needs to learn next.
When the industry reaches this point, "million hours" is no longer the finish line, nor even the most important metric.
What holds real value isn't adding another hour to the data warehouse.
It is that this specific hour happens to fill the exact gap of what the robot previously could not do.
