Robots are held back by the cost of acting in the real world, not by algorithms. A learned world model should sit between agent and world as the main training ground and test bench, with real-world contact kept for data
The Physical AI Voices by Herbert Scale Experts.