Sinolink: The data collection industry is in an early exploratory stage, centered on boundary adjustments and asset upgrades.

Sinolink released a research report stating that data collection services are the segment of the entire data collection industry with the highest order visibility and the most stable cash flow, and leading companies will continue to benefit from the widening supply-demand gap and the industry consolidation brought about by route convergence. In the short term, focus on data collection service providers with order, capacity, and delivery capabilities; in the medium to long term, prefer companies that can platformize data processing and extend into model evaluation, post-training, and deployment feedback. Focus on three types of companies: (1) traditional leaders with experience in traditional AI data services; (2) simulation leaders that build integrated software-hardware platforms based on simulation assets; (3) full-process platforms that layout a closed loop of "datasimulationtrainingmigrationscenario." Sinolink's main views are as follows: The data bottleneck of physical AI is not only a quantity gap, but also a gap in quality and closed-loop capability (1) Quantity: The industry estimates that a mature embodied large model requires at least tens of millions of hours of data, and further progress toward general intelligence may require data on the scale of hundreds of millions of hours, while the industry currently has only hundreds of thousands to one million hours of high-quality training data, leaving a supply-demand gap of about 90%. (2) Quality: Physical AI needs to establish temporal correlations among task instructions, environmental observations, Siasun Robot&Automation states, control actions, and execution results. Data must include not only vision and language, but also information such as joint poses, force perception, and touch, and meet the requirements of "accuracy, completeness, breadth, and authenticity." (3) Interaction: Embodied data collection requires collaboration between edge-side collection and cloud-side processing to form a closed loop of "data collectioncleaning and alignmentannotation and processingquality evaluationdelivery feedback." From an industry profile perspective, the data collection industry is in an early stage characterized by contending technical paths and business models still being explored The divergence among real-machine, body-free, and simulation routes is essentially a balance between precision and scale. Real machines have the highest precision, body-free approaches have the fastest scale-up speed, and simulation has the broadest development potential. The report expects mixed data types to become the mainstream training method for embodied models. The supply gap and unconverged technologies have prompted many players to enter. From a business paradigm perspective, the business models of data collection equipment and services are the clearest, while data assets and platform SaaS have greater long-term potential. The competitive factors of data collection services center on boundary adjustments and asset upgrades (1) Improvements in foundation model capabilities are both a shock and a redefinition of boundaries for embodied data companies. One of the market's biggest disagreements about embodied data companies is whether independent embodied data companies will be crushed by dimensional reduction if general models continue to improve. The report believes that the capability circle of general models lies in "understanding instructions, decomposing tasks, and invoking skills," while precision-task capabilities remain insufficient. GPT's success rate in the precision operation group was only 4.0%, while the control group DM0.5 was 16.75%. Embodied data companies possess rich physical information assets, and their core barrier is reflected in the additional marginal gain in the company's data and services after switching to stronger foundation models. (2) What data companies sell is not video but training assets. Rich physical information, complete data processing capabilities, and continuously optimized training feedback are the core of building data assets. The generalization performance of training data under unseen environments roughly follows a power-law relationship with diversity and strategy. Taking the DROID project as an example, after adding multi-scenario external data, the model's performance within the split and in unseen scenarios improved by an average of 22PP and 35PP, respectively, and generalization capability was significantly enhanced. (3) In the long term, the long-term moat of data collection service companies comes from scenario entry points, evaluation closed loops, and reuse capabilities. The most valuable companies in the medium to long term are often those that control deployment networks or are deeply embedded in deployment processes. Their data sources continue to be generated as business operations proceed, making it difficult for competitors to replicate through one-time capital expenditure. Owning a scenario does not necessarily mean owning data value. It is also necessary to establish a complete closed loop of "failure diagnosisdata recipecollection and processingtraining feedbackreal-machine retesting" to truly achieve model capability gains. At the same time, the reusability of data after migration to different bodies is an important guarantee of data companies' profitability after competition becomes homogeneous. Risk warning Siasun Robot&Automation development progress falls short of expectations, drastic changes in data collection technology routes, and the main businesses of various companies develop below expectations.