Physical AI training data faces an order-of-magnitude gap; the key is turning human actions into data robots can use

Why does robot training need more real-world data?

According to institutional estimates, embodied foundation models need at least tens of millions of hours of real-world data to reach their own “ChatGPT moment.” Currently, the high-quality real-world data available globally amounts to only hundreds of thousands to millions of hours, leaving an order-of-magnitude gap between supply and demand.
As model companies enter physical AI, training data will need to be sourced externally or developed jointly with robot manufacturers. Data collection could therefore generate external demand that can be priced independently.

Agibot’s cross-embodiment data engine, “Chorus众行,” can convert videos of human operations captured by various devices into pre-training data suited to different robot platforms. Tests show it can increase effective robot training data by more than 40 times.
Huayuan Securities believes that obtaining large volumes of first-person data at low cost is only the first step; more important is converting human operations into action training data robots can use. Providers with 3D spatial reconstruction, motion trajectory extraction and physics simulation capabilities have an opportunity to extend from data collection into data assets and training infrastructure.

Tianqi provides embodied intelligence solutions for industrial settings, covering complete robot system adaptation, data collection using physical robots in real industrial environments and intelligent training, as well as customized algorithms for specific application areas.
LUSTER’s optical motion capture systems are already used for humanoid robot data collection and mass-production inspection, and the company collaborates with robot manufacturers and data collection centers. Its optical motion capture system revenue grew rapidly in 2025, driven by demand for embodied intelligence, marking initial commercialization.