Artificial intelligence solves table hockey.

Through simulations

Three students from the University of British Columbia managed to develop an artificial intelligence capable of mastering table hockey, trained exclusively in a virtual simulation. This achievement represents a further step and one of the greatest ambitions of modern robotics, to teach complex machines without spending years on physical tests.

Instead of training the machine and reality, they made it make all the mistakes inside the computer. The biggest challenge of table hockey is that the puck moves at high speed and reacts to small imperfections that are practically impossible to foresee. To solve it, the students used a technique called domain randomization. The idea was not to create a perfect simulation, but to intentionally introduce defects, so the AI ​​trained on crooked virtual tables with different frictions, camera delays and even small variations in the response of the motors.

Faced with millions of unpredictable scenarios, the AI ​​stopped searching for an exact formula and began acting more like a human player, anticipating likely regions where the puck might appear. To further speed up learning, the team used a reinforcement learning algorithm known as Soft Actor Critic. The machine played millions of games against itself at high speed, each defense generated a reward, while each goal received served as a punishment to adjust the strategy.

When the training was finished, the researchers simply transferred software to the physical robot, the impact of this achievement goes far beyond games, demonstrating that an artificial intelligence can learn completely in the virtual world and function immediately in reality, opens the way to accelerate the development of autonomous cars, industrial theft and even rescue drones.