Real robots fail under rain, smoke, glare, occlusion, sensor drift, map staleness, moving targets, unstable contact, and ordinary mechanical wear. A serious system cannot wait for the physical world to supply every rare failure case at full cost.
The practical path is not to replace the real world with simulation. It is to use simulation as a disciplined failure laboratory: generate adverse cases, attach dense labels, replay interventions, record assumptions, and only then ask which claims survive contact with real sensors and real constraints.