The cost of learning from the real world
A physical robot is slow, expensive, and unforgiving to learn from. Every trial takes as long as it takes in real time, there's no pausing to inspect a failure mode from six angles, and a bad control policy can snap a gearbox, crash a drone, or injure someone standing nearby. Hardware wears out from repeated stress testing, batteries need recharging between runs, and a research team is lucky to get a handful of physical robots to share across a dozen experiments. None of this is a minor inconvenience, it's a hard ceiling on how fast anyone can iterate.
Simulation removes nearly all of that friction at once. A simulated robot can fall over, collide, or be driven off a cliff edge thousands of times with no consequence beyond a reset command. Because a simulator is just software, it can also run faster than real time and can run many independent copies in parallel across a cluster of machines, turning what would be months of physical trials into hours of simulated ones. This is why simulation isn't a nice-to-have step before hardware testing, it's the substrate that makes iterating on control algorithms, collecting large volumes of training data, and exploring hardware design choices practical at all.