Video: Robot masters air hockey, learns to beat humans without ever touching real table

- UBC students just dropped a sim-to-real bombshell. They trained an AI to dominate air hockey using only digital twins, bypassing costly real-world trial-and-error. By injecting chaos—warped tables, latency, voltage dips—via domain randomization, the model learned to predict the unpredictable. Ditching heavy physics engines for soft actor-critic algorithms, it mastered the puck’s chaotic bounces in simulation. The result? A robot that crushes humans straight out of the box, no physical practice needed. This isn’t just about table tennis; it’s the blueprint for training drones and autonomous vehicles faster and safer. We’re moving from perfect simulations to gritty reality. The future of automation is learning in the void before hitting the floor.