Embodied AI Data Collection: Will the 5S Store Model Work, or Is Synthetic Data the Endgame?

4.47 billion in funding raised in one year, with nearly 100 players crowding into the embodied data track. This number reminds me of the frenzy in the AI labeling industry back in 2020—hundreds of labeling companies got funded then, too, only for 90% of their final payments to go uncollected a year later.

Core Judgment: The business model for embodied data collection is currently stuck in the "scissors gap" between data quality and collection costs. The 5S store model seems to lower the barrier to entry, but those who will actually make money selling data won't be the players setting up gloves in retail stores. It will be small teams capable of controlling data noise, ensuring annotation consistency, and achieving scalable reproduction.

Let's look at the 5S store model described in the news: ordinary customers receive a set of grippers, gloves, and head-mounted cameras, undergo training, and collect data while doing housework. The initial rollout involves 10 sets of equipment. The engineering issues with this model are:

• Data Noise Control: Ordinary customers have non-standard collection movements. Arm shaking, deviations in gripper angles, and camera occlusions all generate massive amounts of low-quality frames. In anti-fraud domains, we've dealt with similar issues—if user behavior data isn't validated in real-time, false positive rates can exceed 50%. Without real-time quality feedback mechanisms, less than 10% of the collected embodied data might be valid.
• Annotation Consistency: Different people define the start and end points of an action like "picking up a cup" differently. Some count from when the hand touches the cup body, others from when force is applied. Inconsistent annotations lead to gradient conflicts during model training. In industry, annotation consistency usually needs to be above 95% to guarantee model convergence, whereas crowdsourced annotation in the 5S store model hits a ceiling at around 70%.
• Scenario Coverage: The home environment simulated in retail stores is too monotonous. Robustness in real scenarios requires data including lighting changes, occlusions, and random object arrangements. Data collected via the 5S store model is likely "clean but narrow," resembling early facial recognition datasets—everyone facing forward without occlusion, collapsing as soon as a profile view appears.

Comparing this to synthetic data route players (like NVIDIA's Isaac Sim or domestic startups building simulation engines), the difference in engineering feasibility is even more stark. Here is a comparison table:

Dimension	5S Store Scalable Collection	Synthetic Data + Domain Adaptation
Cost per Data Point	~2-5 RMB (incl. equipment depreciation + labor)	~0.01-0.1 RMB (compute cost)
Data Annotation Quality	Low (70% consistency)	High (99% auto-annotation)
Scenario Diversity	Medium (limited by physical space)	High (infinite random generation)
Privacy Compliance Risk	High (collecting faces, indoor environments)	Low (no real persons)
Model Transfer Effect	Requires extensive fine-tuning on real scenes	Requires solving the sim-to-real gap

Key Numbers: Although the cost per synthetic data point is extremely low, solving the sim-to-real gap requires additional investment. Currently, the best domain adaptation techniques can only achieve a generalization rate of about 85% for simulation data; the remaining 15% must be supplemented with real data. This means the core value of the 5S store model isn't "selling data," but "selling that 15% supplement of real-scene data."

But the problem is, can this 15% of real data support an independent business model? Look at the numbers: 4.47 billion in annual funding, averaging 50 million per company, means nearly 90 companies got paid. Yet, embodied intelligence itself is still in its infancy. Fewer than 10 companies can truly run the full chain (data collection + model training + robot deployment). Most data collection companies are essentially doing "data outsourcing" for these 10 companies.

From a risk control engineer's perspective, this outsourcing model has two fatal flaws:

First, poor data reusability. Different robot manufacturers have different sensor models, control interfaces, and action spaces. Data collected by Company A must be re-annotated and converted for Company B. This is like the old days in anti-fraud, where each bank built its own blacklist database, resulting in inconsistent data formats and exchange costs so high nobody wanted to do it. Ultimately, only top-tier players with the highest degree of standardization (like Tesla's Optimus) can form a data flywheel; everyone else gets the scraps.

Second, the long-tail effect of data. Embodied intelligence lacks not common actions like "picking up a cup," but rare scenarios like "pulling a book out from under something in a tilted drawer." The 5S store model cannot efficiently collect these long-tail data because ordinary customers won't encounter such extreme situations. Simulation environments, however, can pair with random physics engines to directly generate tens of thousands of unexpected scenarios.

Quantified Conclusion: If calculated by "valid data frames," the actual output cost of the 5S store model may be underestimated. Assuming a customer collects for 1 hour, generating 3,600 frames, but 70% are invalid due to shaking, occlusion, or annotation errors, leaving only 1,080 valid frames. At a cost of 5 RMB/frame, the effective cost per valid frame skyrockets to 46 RMB/frame. Even accounting for the extra overhead of domain adaptation, the effective cost per valid frame for synthetic data remains under 0.5 RMB/frame. The gap is nearly 100 times.

Therefore, those who can truly make money selling data should be companies that possess both simulation engine and real-scene collection capabilities. They use synthetic data for 80% of "basic training," use the 5S store model to supplement 15% of "real-world fine-tuning," and rely on customized collection for the last 5% of "extreme scenarios" (e.g., hiring professional athletes to simulate falls). Only this combination can push data costs to an acceptable range.

But here comes the question: When embodied intelligence enters large-scale deployment, will this 5% of extreme scenario data become the critical bottleneck determining model safety? Just like financial anti-fraud...

Original Link: https://www.qbitai.com/2026/07/459262.html