Tests whether frontier-scale pretraining, not just small backbones, is the path to better robot policies.
A paper asks whether frontier-scale vision-language representations improve robot control and scales vision-language-action policies to 552 billion parameters, far larger than typical VLA backbones.