Start here

The competition in one paragraph

A robot policy is shown 1 demonstration of a tabletop task — one recording, no language, nothing else — and then has to do the task from a starting position the demonstration did not begin at. Its weights are frozen; the demonstration is the whole prompt. Every submitted model is measured on the same four kinds of change: the objects move, the scene changes, the target object is swapped, and two tasks are chained into one. Each axis is a success rate; the score is the average of the four. A challenger takes the crown by beating the reigning model's average by +3.0 pts on an identical list of units, and every duel is published in full.

Three words you need

The four axes

Where the crown stands

How to read this

Four short pages, in order, and you will know what the competition is and how to enter.

1. How the ladder works — duels, the average rule, and the baseline that holds the crown at genesis.
2. The four axes — what each kind of change is, and how it is drawn.
3. What you submit — a checkpoint, a config, a pinned revision.
4. Honest measurement — why a copy of the reigning model cannot win, and what is published so you can check our arithmetic.

What we measure, and what we do not

We measure one-demonstration adaptation with frozen weights, under change. That is narrower than "how good is this robot policy", and the narrowness is the point.

We do not measure language following — there is no language. We do not measure how good a policy is after fine-tuning on the task — no weight changes. And we do not measure performance on the exact scenes the public demonstrations were recorded in: every unit is perturbed away from them along one of the four axes, so a score cannot be earned by replaying a memorised scene.

Three things that make this different

Every result is published in full. Per-unit outcomes for both sides of every duel, on every axis, with three clips per unit: the demonstration the model was shown, the reigning model's attempt, and the challenger's attempt.

The rule is arithmetic you can redo. The duel page recomputes both sides' scores from the published unit list and shows the comparison line by line. If the page's arithmetic ever disagreed with the published verdict, the page would say so.

Nothing is held back. No private phase, no finalist round, no results withheld. A duel the reigning model wins is published in exactly the same detail as one it loses.