cairn-departure-return

Departure -> return trajectories with exact ground truth, for measuring whether a video world model keeps a world when the camera looks away.

Each episode: a room of well-separated but deliberately confusable objects (two near-duplicate colour pairs, repeated shapes), and a camera that observes a target object, turns away for a measured number of frames, and comes back from a different viewpoint -- so a model cannot pass by replaying its last frame.

A third of the episodes carry a world edit (move, remove, set_attr) scheduled to fire while the target is off screen, for measuring instruction compliance on objects the model cannot see at the moment of the command.

Fields

field	meaning
episode_id	stable identifier
seed, n_objects	scene construction parameters
objects	ground truth per object: id, name, position, yaw, 8-D appearance
cameras	per-frame eye position, yaw, pitch, horizontal fov
visibility	(frames x objects) boolean, rasterised, occlusion included
target_id	the object the camera leaves and returns to
departure_frame, return_frame, absence_frames	measured, not requested
edit	scheduled edit, or null

Metrics

Use cairn.metrics and cairn.runner.run_condition, or score your own model: compare the target's pose/appearance in the frames after return_frame against the frames before departure_frame, and report the result as a function of absence_frames. A method with implicit memory will show a curve that degrades with the x-axis; that degradation is the point of the dataset.

Generated by scripts/build_bench.py in the Cairn repository. Interactive demo: NagaYu/cairn.