Ngseo/stage3-500k-grid-videos

Stage3 500k-bs72 — LIBERO-plus Single-Task Success/Failure Grid

36 model-rollout videos + 1 LIBERO-original demo replay from the Qwen2.5-VL-3B Stage 3 disentangle-infonce 500k-bs72 VLA, all on the same base task: open_the_middle_drawer_of_the_cabinet (libero_goal).

Layout

• 36 model rollouts in the root: 6 perturbation categories × (3 successes + 3 failures)
• 1 reference demo under demos/: a replay of demo_0 from the LIBERO-original open_the_middle_drawer_of_the_cabinet_demo.hdf5 (from yifengzhu-hf/LIBERO-datasets)

We dropped Language Instructions from the grid because the model effectively collapses on language perturbations (only ~1 success across all lang variants for any single base task), so a 3/3 split is not representative for that cell.

Variant selection policy

For each cell we pick variants that are (1) at the highest available difficulty (preferring the hardest perturbations the model can still solve, or fails at) and (2) spread across the available variant-parameter space (e.g. distinct table textures / light setups / robot init states), so the 6 clips within a cell are visibly different scenes rather than near-duplicates.

File naming

Model rollouts: <idx>_<category>_d<difficulty><tag>.mp4

• <difficulty> is the LIBERO-plus difficulty_level (1 easy → 5 hard, or None if metadata absent)
• <tag> is _okexp (3 of 6 per cell) or _failexp (3 of 6 per cell), based on the prior full LIBERO-plus eval. Re-rolling under the same seed reproduces the original outcome for most cells; one Light Conditions variant flipped this run (2 actual OK / 4 actual FAIL vs. expected 3/3).

LIBERO-original demo: demos/demo_open_the_middle_drawer_of_the_cabinet.mp4 (agentview_rgb of demo_0 in the canonical _demo.hdf5).

Reproducibility

• Model: stage3-dual-disentangle-infonce-500k-bs72 (epoch_2.pt, global_step=100k)
• Suite: libero_goal, base: open_the_middle_drawer_of_the_cabinet
• Open-loop chunk: 8 actions; warmup: 10 dummy steps; deterministic
• Camera: agentview, 256x256, 20 fps; flipped to natural orientation

results.json carries per-clip metadata (variant, category, difficulty, expected_success, actual success, steps, frame count, wall time). demos/demo_results.json carries the demo metadata.