Guosen: Embodied data is the core asset of physical AI; model iteration fosters the data industry market.

Guosen released a research report stating that data is the key resource for the development of physical AI large models and the underlying DRIVE supporting model training, capability emergence, and intelligent evolution. Current embodied data accumulation is primarily visual information; human manipulation videos provide rich behavioral and semantic information, while tactile data requires synchronous recording of contact states during interaction, and the supply of existing tactile data remains relatively limited. Embodied data collection routes are diverse, and the training value of multi-source data is prominent. The embodied intelligence data industry chain covers infrastructure, data collection, data processing, and applications. Enterprises with software-hardware collaboration, real-world scenario resources, and continuous data delivery capabilities are expected to form differentiated advantages. Guosen's main viewpoints are as follows: High-quality physical interaction data is key to achieving generalization in embodied intelligence, and the embodied Scaling Law has begun to emerge Looking at the development history of digital AI large models, breakthroughs in model capability boundaries have always been accompanied by the coordinated evolution of data, computing power, and algorithms. Data is the key resource for the development of physical AI large models and the underlying DRIVE supporting model training, capability emergence, and intelligent evolution. The Scaling Law in the embodied intelligence field has begun to receive preliminary empirical support. For example, in experiments with Ego first-person human video pre-training data, Dyna-2 found that when data scale expanded from thousands of hours to millions of hours, model performance showed a clear trend of improvement, providing preliminary evidence for the embodied Scaling Law. Data demand is expected to continue expanding. General model vendors are accelerating their entry into embodied intelligence, and tactile data is highly scarce ARR growth in the digital AI business of some foundation model vendors has slowed, and extension into the physical world is expected to become an important direction for expanding their application boundaries, with demand for embodied data procurement growing rapidly. Multimodal general models represented by GPT-6 Astra continue to expand into the physical world, demonstrating strong task understanding, planning, and closed-loop error correction capabilities in the embodied field, further expanding the entities demanding data. Current embodied data accumulation is primarily visual information; human manipulation videos provide rich behavioral and semantic information, while tactile data requires synchronous recording of contact states during interaction, and the supply of existing tactile data remains relatively limited. Embodied data collection routes are diverse, and the training value of multi-source data is prominent Embodied data collection technology routes include internet video data, Ego first-person video data, simulation synthetic data, and real data, among other types, and data from different sources has complementary value. 1) Internet video data is massive and serves as the data foundation for embodied intelligence, but lacks force and tactile information; 2) Ego first-person perspective data is collected by collectors wearing head-mounted or neck-mounted devices, naturally completing tasks in daily work environments without relying on a specific Siasun Robot&Automation body. It is a route with high scaling potential and low deployment barriers. Unlabeled Ego data is priced at about tens of yuan/hour, processed high-quality data at hundreds of yuan/hour, and Ego devices at the thousand-yuan level; 3) UMI universal manipulation interface data uses a UMI gripper held by the operator as a collection tool to collect end-effector manipulation information and provide structured action information. Its main advantage lies in cross-embodiment reuse without relying on specific Siasun Robot&Automation body hardware, but the action space differs from that of real Siasun Robot&Automation. Data prices range from tens to hundreds of yuan/hour, with core equipment at the ten-thousand-yuan level; 4) Teleoperation data provides real Siasun Robot&Automation execution demonstrations and is a high-value data type, mainly used for model post-training. Data prices range from hundreds to over a thousand yuan/hour, and tactile glove prices vary significantly, with domestic vendors' single-set prices at the ten-thousand-yuan level; 5) Simulation data supports efficient generation and scenario expansion, but core challenges include the Sim-to-real Gap. Domestic vendors are actively deploying data collection and processing, and embodied data is expected to achieve a market space of RMB10 billion The embodied intelligence data industry chain covers infrastructure, data collection, data processing, and applications. Enterprises with software-hardware collaboration, real-world scenario resources, and continuous data delivery capabilities are expected to form differentiated advantages. In terms of market space, looking ahead to next year, industry demand is moving toward the hundreds-of-millions-of-hours level. Calculated at hundreds of yuan/hour, the data market corresponds to potential market space of about RMB10 billion; in the long term, drawing on Tesla's intelligent driving data accumulation process, under the assumptions that vehicles remain continuously in use and each vehicle travels an average of 1 hour per day, by 2024, when the FSD V13 model capabilities achieved a clear breakthrough, Tesla's global fleet driving time is estimated at about 6 billion hours (estimated value, for reference only). Theoretically, the data complexity of Siasun Robot&Automation is far higher than that of intelligent driving vehicles. If embodied intelligence achieves a major model capability breakthrough similar to FSD V13, data demand will likely exceed that duration. Considering future declines in data collection costs, the long-term data market size is also expected to exceed RMB100 billion, indicating broad market space. Risk warnings: Embodied intelligence model progress falling short of expectations, embodied Siasun Robot&Automation shipments falling short of expectations, embodied data procurement demand lower than expected, changes in data technology routes, data price declines, and intensifying industry competition.