Beyond LLMs: Why Video Games are the Secret to Physical AI Common Sense
Researchers are moving away from LLMs for physical intelligence, instead utilizing high-fidelity video game environments to train agents in complex spatial reasoning and physics-based interactions.
The quest for Artificial General Intelligence (AGI) is hitting a wall with Large Language Models (LLMs). While models like GPT-4 excel at processing text and code, they lack a fundamental understanding of the physical world—a gap that "Physical AI" aims to bridge. Industry experts are now arguing that the internet's text-heavy data is insufficient for robots to understand cause and effect in the real world. Instead, a new wave of startups is turning to video games as the ultimate training ground.
By leveraging the sophisticated physics engines of modern gaming, developers can expose AI agents to millions of hours of simulated physical interactions. These environments provide high-fidelity data on gravity, friction, and spatial navigation that the flat text of the internet cannot offer. This approach allows AI to learn "common sense" physics—such as knowing a glass will shatter if dropped—before it ever enters a physical robotic body. The shift suggests that the next breakthrough in AI won't come from better chat scripts, but from agents that have "lived" in complex, simulated 3D worlds.
Source: TechCrunch
