Human Demo Inputs — LAPA source footage
Real factory recordings of a human performing the paper-into-feeder task from an ego-POV. Source material for the v2v_sdedit experiments (sessions 2–3).
Original human-ego recording. Full-length demonstration of the cart→feeder paper-handling action at native capture rate. Source-of-truth for human-demo data.
▶ open file directly
Same content as ego_shaun.mov, downsampled to cosmos's native 16 fps.
This is the file v2v_sdedit consumed in later sessions — feeding 30 fps caused
silent truncation to ~93 source frames, dropping the back half of the action.
GR1-Rendered Outputs — cosmos-predict2 inference
robot_ego_v2.mp4 matches that geometry. All other outputs were generated
in 3:4 portrait (480 × 640), which is off-distribution.
GR1 ego-POV clip in cosmos's native training spec (16:9 landscape, 432 H × 768 W, 93 frames at 16 fps). The only video on disk that already matches the on-distribution geometry. Likely sim-rendered. Format target for Phase 1 sim data.
▶ open file directly
video2world_gr00t output, seed 42, guidance 8, on
fourier_prox_feeder.png (GR1 close-up near the press's angled feed deck).
Session 5 "best in proximity" experiment. Result: pivot + slight angle but never placed.
video2world_gr00t output, seed 42, on fourier_worker.png
(wide-shot: GR1, paper, press from a distance). Session 4 3-seed sweep. Result: GR1
pivots toward feeder, paper occludes deck during pivot, never lowers paper onto it.
video2world_gr00t output, seed 42, on fourier_over.png
(GR1 near press front cabinet, no visible feed deck). Session 5. Result: paper placed
on the flat cabinet surface — wrong target — because the angled deck wasn't visible.