Enter the competition

One competition. Two tracks. Submit once.

Enter a model in the open simulation qualifier and iterate on the live public board. The top teams advance to the real-robot final, revealed on stage at the Nebius Physical AI Summit in November.

Company track

Startups & product teams

Bring a shipping model and see where your product actually stands against the field – an objective, third-party read on identical tasks, not a self-reported number.

  • Submit a shipping model as an endpoint or container
  • Benchmark against baselines and peers on identical tasks
  • Advance to the real-robot final on merit
Prize details announced with the competition.
Researcher track

Labs & individual researchers

A common yardstick does the most here – identical tasks let otherwise incomparable submissions rank directly against each other.

  • Enter a lab model or a fine-tune of a public checkpoint
  • Iterate against a live public leaderboard through the qualifier
  • Finalists run on real Franka + Robotiq hardware in November
Compute credits and a place on the summit stage for finalists – exact rewards confirmed with the competition.

What you submit

Submit once – the same endpoint runs in both simulators and on the real robots.

Your model runs as a server that speaks one open protocol. Our simulator – and, in the final, the real robot – connect to it as a client: they stream observations to your model and execute the action trajectory it returns. One interface, every target.

Option A

Hosted inference endpoint

A webserver that exposes the WebSocket endpoint. Well suited to models that load fast and serve over the wire.

Option B

Docker container

Package your model with its dependencies. Handy for heavy models – it streams a loading status so the connection never times out while weights load.

One open WebSocket protocol, msgpack with numpy. The client streams an observation each timestep – camera images, robot state, the language task – and your model replies with a short action trajectory. Use the keys your model needs; ignore the rest.

Read the connect-your-model docs →
Task suite

Standardized manipulation tasks

The qualifier runs across two independent simulators. The real-robot final broadens onto DROID and PhAIL pick-and-place on Franka + Robotiq hardware.

Qualifier · simulation

Real-robot final · Franka + Robotiq hardware

Rules essentials

How it stays un-gameable

  • 1The public board runs through the summer as teams iterate – seeded from day one with baseline models.
  • 2Finalists are selected on a held-out evaluation set – unseen initial states, scene variations and task variants – scored once when the qualifier closes.
  • 3Models are frozen one month before the final deadline.
  • 4The final's decisive ranking is scored on held-out conditions the teams never saw.
  • 5Baselines run in both phases – the bar to beat, and a check that the field is really improving.
See the live leaderboard

Concept preview – registration and exact dates are set with Nebius when the competition is announced.