Humanoid Reality Checks, VLA Data Hunger, and the Sim2Real Money Problem

Scaling Laws for Vision-Language-Action Models

Finally someone ran the numbers on how much data VLAs actually need before performance plateaus, and the answer is more than most teams are collecting. If you're bootstrapping a manipulation dataset, this tells you where the diminishing returns kick in so you stop burning compute chasing marginal gains.

Cheap Sim-to-Real Transfer Without Domain Randomization Overkill

Domain randomization has become a crutch that eats sim time without buying much real-world robustness. This paper shows a leaner recipe that gets similar transfer results with a fraction of the randomization budget, which matters if you're paying for cloud sim hours.