The contest

The competition asks one question. A policy is shown 1 of a task and none, then has to do that task. What the demonstration contains and where the policy is then put are the two facts that decide what a score means.

The contest Vector: one demonstration from another scene, level 1

The policy is shown one demonstration of a task - the frames of three cameras, the action trajectory and the proprioception - recorded by RoboTwin's expert in one scene, and must do the task in another scene of the same level. Level 1 varies the objects' positions and yaw between the two. Submissions are weights of the BPP architecture only.

The policy is shown the frames, the action trajectory and the proprioception — everything the robot saw and everything it did — and is then scored from a different initial state. Its skills are Sixteen tabletop tasks (two Franka arms, level 1), and a duel runs 160 per side.

Replaying the demonstration

Can the demonstration be replayed? no — the actions do not fit a different initial state. The rollout is scored from a different initial state, so what the contest measures is how well a policy carries one demonstration over to a scene it was not recorded in, not how faithfully it plays one back.

Honest measurement sets out what that does and does not let a score tell you.

What is fixed

One demonstration, no language, both sides of a duel on an identical unit list with identical demonstration bytes, the crown rule at +3.0 pts, and every record published in full — per-unit outcomes, both sides, and the clips.

Entering

A submission is a Hugging Face repository at a commit sha, holding not this contest: submissions are weights, and nothing of them runs and the policy code it names, queued against the contest and run in a sandbox with no network — the same policy on every skill. See What you submit.