Frequently asked questions

You don't need either. The embodiment, the task set, the evaluation environment, the base model, the tele-op dataset, the overnight H100 allocation and on-site GB10 hardware are all provided. Bring a laptop and an opinion about action chunking.

That's the diagnosis, not a disqualification — it's also the name of the event. Sim rounds get you onto the board; the final is on real hardware. Most of the useful learning happens in the distance between the two.

You post-train or fine-tune a small pre-trained manipulation model. Around 3B parameters is the sweet spot for the compute budget, but you're free to optimise a larger open-source model if you think you can make 2–3 runs on H100-class hardware stretch that far.

Open-source base models, yes. Synthetic data you generate during the event, very much yes — it's one of the more interesting axes to compete on. Private datasets or a policy you trained last month, no.

Twenty teams, eighty builders — so four is the shape. Apply solo and we'll match you at kick-off; a lot of the best pairings at these things are made in the first hour.

No. Free if you're accepted — compute, data, bench, food and floor time are covered. Applications are read for what you'd do with the compute, not your CV.

Objectively and automatically, across multiple tasks on one embodiment — no demo-day charisma vote. A separate bounty panel awards side prizes for approach and data strategy, which is where human judgement belongs.

Top teams on the virtual leaderboard qualify for Sunday afternoon's live final. Our operators load your selected policy checkpoint onto the physical robot arm evaluation bay.

It stops every robot on this site. DO NOT SPAM CLICK IT.