The main robot “brains” (VLA modelsVLA (Vision-Language-Action)A robot “brain” that looks (camera), reads an instruction and directly decides which movements to make.) compared across the 7 capabilities of an autonomous robot. For each capability, we use the model's best published success rateSuccess rateOut of 100 attempts, how many times the robot completed the task. on the same reference test. “Not measured” means no paper has published a result: it is not a zero.
Choose up to 3 models:
For a like-for-like comparison, each capability is measured on the same reference test for every model. End of the gray bar = ideal robot (100%); vertical line = current record on this test. “≈”: reference test missing, average of the capability's other tests.
Physical Intelligence · 2025 · ≈3 billion parameters
The most versatile of the open models: designed to work in homes it has never seen.
On real robots (RoboArena), wins about 32% of its head-to-head matchups against the top-ranked model on the official leaderboard.
Stanford · 2025 · 7 billion parameters
OpenVLA retrained with a better recipe: much more accurate and about 25 times faster.
NVIDIA · 2025
NVIDIA's model designed for humanoid robots, trained at scale in simulation.