Physical AI’s data bottleneck: Why might purchases of collection hardware scale up before robots?
The material identifies high-quality training data as an industrial bottleneck for physical AI, with opportunities concentrated in collection hardware, real-world data and simulation services. Estimates point to steep growth in market size, but companies are at markedly different stages of realization, from having made sales or accumulated data to having disclosed no revenue yet.
Physical AI’s demand for data brings purchasing opportunities in collection equipment and sensors forward, but whether companies benefit depends on specific supply relationships, order growth and the realization of data services revenue.
Why is research shifting toward data collection?
The material shifts the central bottleneck for physical AI from robots themselves to data: deploying large models may require tens of billions of hours of high-quality data, while existing open-source datasets are only on the scale of millions of records and hours. The scale and quality of data supply therefore constrain the expansion of training capacity.
An institution estimates that global demand for robot data collection will rise from approximately 10.8 million hours in 2027 to 2.925 billion hours in 2031, with the corresponding market expanding from 7.8 billion yuan to 824 billion yuan. This projection depends on continued growth in training demand and a matching expansion in collection, labeling and processing capacity; the material notes that labeling capacity and high-precision collection hardware remain scarce.
What problems do low-cost data and real-robot data respectively solve?
The four collection paths are real-robot teleoperation, robot-free collection using UMI and Ego, simulation and synthesis, and world-model generation. The robot-free route costs less and is easier to deploy in advance, while real-robot data serves as the “gold standard” for training validation; the two serve different stages.
Cost differences determine the data mix: Ego visual data is quoted at approximately 120–500 yuan per hour, while real-robot and teleoperation data costs approximately 500–2000 yuan per hour; simulation and synthetic data provide a low-cost supplement. The institution expects Ego visual data to become the largest category, followed by a scaling up of tactile and real-robot data.
Technical architectures also have distinct roles. The material judges that general-purpose large models handle planning and semantic understanding, VLA converts visual and language information into action control, and world models predict the consequences of actions. Fine manipulation and contact coordination still require targeted control training, so data demand spans multiple dimensions, including vision, pose and touch.
Which segments could see purchases materialize earlier?
Industry research suggests the following sequence: if orders for data-collection equipment manufacturers scale up in 2027, they will need to purchase IMUs in bulk beforehand. IMUs sense motion, pose and spatial trajectories, so sensor purchases may precede delivery of end equipment; the key condition is that equipment orders expand on schedule.
Orbbec launched a robot-free data-collection hardware platform in July 2026, covering first-person perspectives, handheld operations, wrist interactions and device synchronization. Its customers include data-collection service providers, large-model developers and robotics companies. Its business connection lies in hardware purchases made by these customers when they build collection systems.
A case closer to the sales stage is Skyworth Digital: the material cites the company’s investor communications as saying that its EGO-series data-collection products have generated sales both domestically and overseas. Product launches, customer onboarding and actual sales are separate stages, and assessments of company progress should be tied to the specific stage.
Data has already accumulated, but how far away is revenue?
The industrial data-collection and practical training center operated by Tianqi has accumulated nearly 20 thousand hours of multimodal data, deployed more than a hundred robots and covered automotive, battery and 3C manufacturing scenarios. Its connection lies in combining real industrial settings, collection equipment and labeling management systems to produce training datasets.
Commercialization in simulation is at an earlier stage. Fengyuzhu has entered physical simulation, Sim2Real training and spatial synthetic data through a controlled joint venture, but has disclosed no revenue yet from its physical AI business; the material directly ties its business upside to orders materializing at Fengying Technology. Between industry market expansion and revenue growth at an individual company lies the additional step of converting opportunities into orders.
