Why does robot training require data collection?

  • Robots use sensors, cameras, radar, and other equipment, combined with human operation or simulation, to collect multidimensional raw data from real-world scenarios as material for subsequent AI training.
  • Data must cover multiple scenarios while also balancing accuracy, compatibility, and efficiency at scale. As general-purpose foundation models advance, the report forecasts growth in data demand, to be met through both external procurement and companies’ own collection.

What are the trade-offs among the four collection approaches?

  • Teleoperation provides the highest-quality data but is more costly and is mainly used for high-quality imitation learning data; UMI balances cost and quality and can quickly collect demonstration data when physical robots are difficult to deploy at scale.
  • EGO is currently the mainstream approach, with lower costs, and is mainly used for foundation model pretraining. Simulation data theoretically offers the highest production capacity and lowest cost, with considerable potential for large-scale pretraining.
  • The report expects multiple technical approaches to be used in different scenarios in the future: trade-offs among collection quality, cost, and speed of expansion determine their respective suitable uses.

What are the midstream market opportunity and industry developments?

  • Wen Ji of Zheshang Securities estimates under a neutral scenario that the midstream data collection market will reach 174.1 billion yuan by 2030, with a compound annual growth rate of 203% from 2026 to 2030. The firm believes this segment is in an early development phase, moving from 0 to 1, and says dataset suppliers, integrated software-and-hardware collection providers, data post-processing firms, and simulation platforms warrant attention.
  • The Hangzhou National Artificial Intelligence Application Pilot Base (Embodied AI) is undertaking a capital increase totaling 200 million yuan, with registered capital proposed to rise from 100 million yuan to 300 million yuan; the capital increase and shareholding restructuring are expected to be completed by the end of 2026.
  • Beijing Noitom, a subsidiary of Leyard, has served multiple robot manufacturers and participated in building several robot training and data collection centers, providing solutions that include motion capture equipment, data collection, and scenario applications.