Towards physical AI.

From the laboratory to the production line

Today, training autonomous robberies still depends on slow, expensive and limited data collection. In addition, artificial intelligence policies that work well in simulation frequently fail when transferred to real robberies. This discrepancy slows down innovation in sectors such as manufacturing, health, logistics and agriculture.

To overcome that problem, Lightwheel developed its simulation platform built on Nvidia systems such as Omniverse, Isaac SIM and Isaac Lab. Data generation occurs on a large scale, operators use virtual reality headsets, spatial control devices and exoskeleton systems to command Simulated humanoid robots in complex industrial tasks.

That teleoperated data is then amplified by techniques that increase its value up to 1000 times, creating robust and diversified training sets, the GR00T fundamental model is fine-tuned with that synthetic data, combining images, joint states, task descriptions and scene metadata.

The mind exists before the body.

The entire process goes through a two-stage quality assurance, combining automated validation and manual review to ensure physical realism and consistent behavior. These models were implemented in the Unitree G1 humanoid robots at the Gilly automotive factory. The robots perform transportation tasks, precise positioning of components and coordinated manipulation of heavy loads, maintaining stability in dynamic environments with humans around, with development cycles reduced from months to weeks and a ratio of 100 to 1 between simulated and real, the platform demonstrates that simulation is no longer an experimental support and that perhaps the revolution is not in robots learning faster, but in the fact that they now learn before even physically existing.