We ran 1,024 domain-randomized worlds against a single hand-tuned scene. The randomized policy lost in simulation and won on hardware. Here's the data.
We took a single warehouse-pick task and trained two policies. The first saw one carefully hand-tuned scene; the second saw 1,024 domain-randomized variants — friction, mass, lighting and clutter all perturbed per world. In simulation, the hand-tuned policy scored higher. On real hardware, it fell apart.
The randomized policy lost in sim because it refused to overfit the easy case. That refusal is exactly what transferred: on the physical arm it completed 91% of picks versus 38% for the hand-tuned baseline. The lesson is not subtle — generalization is bought with variance, and variance is bought with GPU-seconds.
The cost to run those 1,024 worlds on the network was $2.41. On a hyperscaler at on-demand H100 rates, the same job would have been roughly six dollars and a queue. When rehearsals are cheap, you stop rationing them, and the sim2real gap stops being a research problem and starts being a budgeting one.
We're publishing the seeds and the scene definitions so anyone can reproduce the runs in SimArena. Volume, not heroics.