Technology

In 2018, a simulated humanoid learned to do a backflip by watching a motion capture clip. In 2025, the same lineage of research produced a single model that controls a real Unitree G1 robot with zero task-specific training. The path from one to the other redefines what ’learning to move’ means for robots, virtual humans, and the film industry.

In a Google DeepMind lab, a robot arm folds paper into a fox shape from fewer than 100 demonstrations. At ICLR 2026, 164 papers compete to define robot intelligence. From RT-1 to BitVLA, the science behind Physical AI is moving faster than the machines themselves.

A CMU roboticist says the hardware is solved. The real challenge is building a brain that understands your body. From experiment-free RL to neural interfaces, the science of wearable Physical AI is converging, and so is the market.