Physical AI data collection: How does commercialization differ from perception hardware to the training loop?

Why is data collection more than just a sensor business?

This review by CICC Machinery divides physical AI data collection into optical perception, mechanical and multimodal perception, motion capture, and closed loops involving data, simulation, and agents. It presents a map of business initiatives: different companies may provide collection tools, training material, or platforms that use data to train robots.
These areas connect to different robot capabilities: vision provides spatial information, touch and force sensing record contact feedback, motion capture provides movements and real-world trajectories, and simulation platforms provide virtual training environments. Thus, even within the data-collection theme, product uses and revenue sources differ markedly.

Orbbec has launched data-collection platforms for first-person perspectives and handheld operations; OPT Machine Vision is expanding from core machine-vision components such as light sources, lenses, and cameras into embodied visual perception. The former already has clearly identified collection-platform products, while the latter represents an extension of its existing components business into new applications. Their stages of development should be considered separately.
Fulaixin Material is developing data-collection gloves and curved fingertip sensors to record hand and gesture data for training physical AI models; Ampron is entering data collection with six-axis force sensors. Their respective points of connection are motion recording and mechanical feedback, and their business relevance depends on the hardware used in specific collection solutions.
OptiTrack, a subsidiary of Leyard, provides optical motion-capture systems for embodied-intelligence motion capture and the collection of kinematic data and real-world trajectories. Its role in this chain is to provide the motion-recording tools needed for training.

Hesai has launched Kosmo, a three-dimensional world data engine that integrates lidar and cameras to provide spatial data and cloud services for physical AI. Its activities extend from perception hardware to data delivery, broadening the business focus from equipment to data and services.
Speechocean has built an embodied-intelligence motion-control data-collection system and lists related customers; Baotong Technology is working with Guanglun Intelligence on data collection and validation in real-world settings including mines, steelworks, and ports. The two enter the field through training-data services and industrial scenarios respectively, and their points of value connection differ.
Simulation is another path: Fantawild Digital Technology is pursuing digital twins and physical AI training, providing the virtual environments needed for robot iteration. Its activities relate to supplying training environments, a different business activity from selling collection hardware.

The table shows that Baosight Software’s embodied-intelligence orders on hand have exceeded 200 million yuan. Compared with product launches or application directions, orders more directly reflect customer purchasing progress; however, orders on hand and recognized revenue are different measures.
Century Hengtong’s overall data-labeling revenue in 2025 exceeded 100 million yuan. This amount covers its entire data-labeling business and cannot be attributed directly to embodied intelligence. Readers should distinguish the scale of existing business from the contribution of a specific business line.
The table shows wide differences in estimated 2026 price-to-earnings ratios: for example, 153 times for Orbbec, 74 times for OPT Machine Vision, and 230 times for Speechocean. These ratios reflect each company’s overall estimated earnings, not a separate valuation of its data-collection business; comparisons should also consider each company’s products, customers, and stage of commercialization.