masterwu/WM4VLA_benchmark_v4

WM4VLA Benchmark v4

Paper-definitive evaluation benchmark for video world models — 240 samples across 6 robotic manipulation datasets (agibot / airoa_moma / droid / interndata / rh20t_cfg5 / rh20t_cfg7), filtered for quality and subsampled via sqrt-N–weighted k-means medoid selection.

This supersedes masterwu/WM4VLA_benchmark (v1, 1000 samples, retired 2026-04-17).

At a glance

Samples	240
Datasets	6 (agibot 77, airoa_moma 35, droid 54, interndata 30, rh20t_cfg5 23, rh20t_cfg7 21)
Frames per sample	81 (aligned so frame 0 passes arm-visibility)
Resolution	480 × 640
FPS	per-sample native (agibot 15, airoa 30, rh20t_cfg5 12, rh20t_cfg7 10, droid 14, interndata 30)
Size on disk	~6.7 GB
License	MIT (see upstream dataset licenses for source terms)

Provenance (v4 = filter-backport of v1)

1. v1 pool (1000 samples): original k-means-selected evaluation benchmark.

2. Filters applied (from the v3 protocol):

   • num_frames ≥ 81 before any trim.
   • Arm-visibility three-tuple on gripper_scenario.mp4:
     • Frame-0 non-background pixel count > 3000 (T0)
     • Min-window pixel count > 500 (T1)
     • Max/min ratio < 3.0 within the 81-frame window (R)
     • Scan up to 200 frames to find the earliest valid window.

3. Re-encoded via ffmpeg with libx264 crf=18 for samples whose valid window starts past frame 0 (trim offset stored in arm_visibility_v4.json).

4. Post-trim sanity re-checks the re-encoded mp4; failures are replaced by next-best survivors to restore per-dataset allocation.

5. k-means subsample to 240 total, allocated by sqrt-N per-dataset with largest-remainder rounding:

   dataset	v1 pool	survived filters	v4
   agibot	490	335	77
   airoa_moma	103	91	35
   droid	247	238	54
   interndata	77	77	30
   rh20t_cfg5	45	45	23
   rh20t_cfg7	38	37	21
   total	1000	823	240

Note: train_manifest.json from v1 is unchanged by this filtering — existing training runs that use the v1 train split are unaffected.

Per-sample layout

{dataset}/{episode}/{camera}/
  ├─ rgb.mp4              # H.264, 480×640, 81 frames, per-dataset FPS
  ├─ gripper_scenario.mp4 # skeleton-rendered gripper/arm projection (for arm-visibility & ROI)
  ├─ caption.pickle       # {"caption": str, ...}
  ├─ episode_meta.npz     # ee_pose (4×4 per frame), joint_angles, gripper_openness, camera intrinsics/extrinsics
  └─ overlay.mp4          # optional debug overlay (rgb + gripper_scenario); present for ~163/240 samples

agibot uses dual-arm keys ee_pose_left / ee_pose_right (+ gripper_openness_left/right + joint_angles_left/right); all other datasets are single-arm and use ee_pose / gripper_openness / joint_angles.

Top-level files

• benchmark_eval_split_v4.json — flat list of 240 samples with {embodiment, dataset, episode, camera, data_root} per entry. Drop-in for baseline adapters (see baseline integration guide).
• benchmark_manifest.json — per-dataset export summary (counts, backfills, trim stats).

Evaluation protocol (recommended)

• Input window: first 81 frames of rgb.mp4 starting at frame 0. v4 guarantees ≥ 81 frames; no freeze-pad needed.
• Metric window: first 49 frames of both GT and generated output (post-generation slice). Matches the Kinema4D convention.
• Skip frame 0 in per-frame PSNR/SSIM/LPIPS — frame 0 is the I2V input image, its metric is trivially perfect.
• Metrics: PSNR ↑, SSIM ↑, LPIPS ↓ (AlexNet backbone), FVD ↓ (I3D Kinetics-400), FID ↓ (InceptionV3), Latent L2 ↓ (Wan2.1 VAE), tLPIPS ↓ (temporal LPIPS between consecutive generated frames, GT-free).
• Reporting levels:
  • PSNR / SSIM / LPIPS / tLPIPS / Latent L2: overall + per-embodiment + per-dataset
  • FVD / FID: overall + per-embodiment only (per-dataset N ≈ 30-80 is below the empirical FID/FVD stability threshold)

Baseline usage

import json
split = json.load(open("benchmark_eval_split_v4.json"))
for s in split["samples"]:
    rgb = f"{s['data_root']}/{s['episode']}/{s['camera']}/rgb.mp4"
    meta = f"{s['data_root']}/{s['episode']}/{s['camera']}/episode_meta.npz"
    # ... run your baseline, save gen.mp4 to
    # outputs/<baseline>/{s['embodiment']}/{s['dataset']}/{episode_basename}/gen.mp4

Citation

@dataset{masterwu_wm4vla_benchmark_v4_2026,
  author = {wuzy2115},
  title  = {WM4VLA Benchmark v4},
  year   = {2026},
  url    = {https://huggingface.co/datasets/masterwu/WM4VLA_benchmark_v4},
}

Changelog

• v4 (2026-04-17): Current. v1 pool filtered by arm-visibility + min_frames ≥ 81, subsampled to 240, re-encoded for trimmed samples, post-trim sanity + backfill to restore per-dataset allocation. Adds episode_meta.npz per sample (was absent in v1 export).
• v1 (2026-04-12): Retired. 1000 samples, unfiltered pool. Archived at masterwu/WM4VLA_benchmark.