---
license: cc-by-4.0
language:
  - en
tags:
  - robotics
  - video
  - hallucination-detection
  - world-models
pretty_name: WMBench — World-Model video benchmark for hallucination detection
---

# WMBench

Benchmark suite for evaluating **hallucination / anomaly detection** on
video generated by world models (Cosmos, OpenSora, etc.) against real
robot training distributions.

## Datasets

| Dataset | Source | Tasks | Videos |
|---|---|---|---|
| `gr-1/`  | NVIDIA GR1 (`PhysicalAI-Robotics-GR00T-GR1`) | 5 eval tasks | 5 real + 24 Cosmos gens |
| `droid/` (planned) | Stanford DROID | — | — |

Each dataset folder contains:

- `training/`        Real training mp4s (one per task)
- `generated/`       World-model generated mp4s, grouped by task
- `reference/`       50 SAM3-segmented PNG frames per task (for null calibration)
- `null_per_task/`   Pre-computed cycle null distributions (.npz)
- `results/`         WarpDyn scores, ranking, visualizations
- `method.md`        Per-dataset method notes / reproducibility

## Method: WarpDyn

Pure feature-matching anomaly detector using RoMa cycle composition error,
per-task multi-lag null distribution, and per-task ratio scoring.

See [`gr-1/method.md`](gr-1/method.md) for the full step-by-step method.