Robot Data Industry Insights: How to choose between real-machine teleoperation, simulation, and human first-person perspective? The 'pyramid' framework mentioned in an interview with @axisrobotics founder is quite interesting: ▪️ Real-machine teleoperation: Most complete physical information, but costs around $200/hour and is hard to reuse across different robot bodies. ▪️ Simulation data: Scalable and controllable, but there's a 'simulation-to-reality' gap issue. ▪️ Human first-person perspective: Low cost and rich scenarios, but the training value heavily depends on the future form of the robot body (humanoid/wheeled/robotic arm...). If robots are to have their own 'GPT moment,' the required data might reach the scale of 100 million hours (compared to GPT-3's ~15 million hours of internet text data). This is also why embodied intelligence data companies in China raised over 17 billion RMB in funding in the first half of 2026.