Millions of attempts, not a handful
Reinforcement learning typically requires an enormous number of trial-and-error attempts to learn a good policy, often millions of individual actions across thousands of episodes. Early in training, the policy is close to random, so most of those attempts are bad ones, the equivalent of a robot arm swinging wildly, missing objects, or a legged robot falling over again and again. That's simply how the learning process works, the agent needs to experience a wide range of outcomes, including plenty of failures, to figure out which actions lead to reward.
Doing that on physical hardware is both impractical and often unsafe. A real robot arm crashing into things millions of times would destroy itself, wear out its motors, or damage its surroundings long before it learned anything useful. Physical trials are also slow, a real robot can only act roughly as fast as the real world allows, while a simulation can run many instances in parallel, faster than real time, on a server rack with no risk to any hardware at all.