Nebius × Positronic Physical AI Competition opens to test robot manipulation models

The competition builds on an existing relationship: Nebius is already a founding partner of Positronic's PhAIL.ai leaderboard, which tracks physical AI performance independently. By adding a contest with simulation rounds and a physical robot arm, the two companies are trying to turn that leaderboard into a more structured proving ground for robotic manipulation models.

Positronic points to a gap in how robotic foundation models are compared. Its DROID benchmark, run on the same Franka robotic arm, tested open models across more than 1,000 episodes covering over 100 everyday manipulation tasks. The spread was wide — the strongest models completed about 70 percent of tasks on average, versus around 12 percent for the weakest — which the organizers read as evidence that manipulation is still an open research problem.

The organizers cite the DARPA Grand Challenge as inspiration, suggesting they see standardized technical challenges as a way to speed progress in physical AI, much as that contest did for autonomous vehicles. Whether a competition format can do the same for robot manipulation is an open question, but the involvement of Nebius's GPU infrastructure and Physical AI Workbench gives participants a concrete stack to build on.