transition-gen-world (world-model format)

RoboCasa coffee3 ground truth (165 episodes) plus 144 world-model-generated transition clips, in the Cosmos Predict 2.5 multiview training layout. The generated clips target gripper and subtask transitions, which occupy under 2% of frames in the GT corpus and are the regime the world model reproduces worst.

videos/{tag}_{left|right|wrist}_{IDX}.mp4
metas/{tag}_{left|right|wrist}_{IDX}.txt
cameras_real/{tag}_{IDX}_{extrinsic|intrinsic}_{head|hand_0|hand_1}.txt
reason1_embeddings.pt
lists/  all_309 | gt_165 | gen_144 | gen_gated_two_labeler
EPISODE_MANIFEST.json

Generated clips use separate tags (press_button_gen, serve_mug_gen, setup_mug_gen) so RC_MV_TAGS can select GT-only or generated-only without rebuilding. 81 frames, 480x480, 3 views; camera files are the source episode's rows sliced to the generated window.

	episodes	note
GT	165	the exact corpus the round-2 world model trained on
generated	144	from Gom-sy/cosmos_predict25_mv_coffee3_r2_9435, 1 conditioning frame, guidance 7, 35 steps

Transition presence measured on the clips themselves (IDM gripper sign flip, not the requested anchor): 85% contain a flip, 79% match the requested direction. Two independent IDM labelers agree on the flip for 101 of 144 clips (median timing difference 1 frame); that subset is lists/gen_gated_two_labeler.txt. The world model places transitions on average +24 frames later than requested, so filtering by the requested offset discards usable clips - filter by the observed flip instead.

Clip-level provenance, intended vs observed anchors, and the setup_mug metadata errata are in Gom-sy/robocasa_coffee3_transition_gen (transition_meta_fixed.json, README errata).