Blame for README.md

---
license: apache-2.0
pretty_name: WorldFailBench
tags:
- robotics
- world-models
- benchmark
---

# WorldFailBench Dataset

This repository contains the open-source dataset for the paper:

> **WorldFailBench: Do World Models Understand Policy Failures?**

WorldFailBench evaluates whether *world models* can serve as faithful policy
evaluators — not only predicting **whether** a robot policy succeeds or fails,
but also **how** it fails (failure reason), **when** it fails (failure stage),
and **under what conditions** it transitions from success to failure (failure
boundary). This dataset provides the **simulator ground truth** used to score
world-model behavior: real closed-loop rollouts executed in the
[RoboTwin](https://robotwin-platform.github.io/) and
[LIBERO](https://libero-project.github.io/) physics simulators, together with
success/failure labels, task-stage annotations, and controlled success/failure
boundary cases.

---

## Dataset Overview

| Subset | Simulator | Contents | Scale |
|---|---|---|---|
| `RoboTwin/` | RoboTwin | First-frame image + metadata per evaluation episode | 5 policy configurations × 20 tasks × 50 episodes = **5,000 episodes** |
| `LIBERO/` | LIBERO (libero_10) | First-frame image (+ raw array) + scene metadata per evaluation episode | 3 policy variants × 10 tasks × 50 episodes = **1,500 episodes** |
| `boundary/` | RoboTwin + LIBERO | Paired success/failure boundary cases built by step-wise perturbation of the initial scene, with full rollout videos | **120 pairs, 2,400 rollout episodes** |

```
dataset/
├── RoboTwin/                     # RoboTwin evaluation episodes
│   ├── pi05/                     # π0.5 policy
│   ├── pi05-data-20/             # π0.5 fine-tuned with 20% of the training data
│   ├── pi05-data-50/             # π0.5 fine-tuned with 50% of the training data
│   ├── fastwam/                  # FastWAM policy
│   ├── lingbot-vla-v2/           # LingBot-VLA v2 policy
│   │   └── <task_name>/          # 20 dual-arm tasks per policy
│   │       ├── episode_000/
│   │       │   ├── first_frame.png
│   │       │   └── metadata.json
│   │       ├── episode_001/
│   │       └── ...               # 50 episodes per task
│   └── ...
├── LIBERO/                       # LIBERO evaluation episodes
│   ├── pi05_data20/              # π0.5 fine-tuned with 20% of the training data
│   ├── pi05_data50/              # π0.5 fine-tuned with 50% of the training data
│   ├── pi05_data100/             # π0.5 fine-tuned with 100% of the training data
│   │   └── task_XX/              # 10 LIBERO-10 tasks per policy variant
│   │       ├── episode_000/
│   │       │   ├── first_rgb.png
│   │       │   ├── first_rgb.npy
│   │       │   └── scene.json
│   │       ├── episode_001/
│   │       └── ...               # 50 episodes per task
│   └── ...
└── boundary/                     # Success/failure boundary cases
    ├── LIBERO/
    │   ├── manifest.json         # LIBERO pair list + selection criteria
    │   ├── orientation/          # pair_001 – pair_020   (20 pairs)
    │   ├── position/             # pair_021 – pair_040   (20 pairs)
    │   └── size/                 # pair_041 – pair_060   (20 pairs)
    └── RoboTwin/
        ├── position/             # pair_001 – pair_020   (20 pairs)
        ├── orientation/          # pair_021 – pair_040   (20 pairs)
        └── size/                 # pair_041 – pair_060   (20 pairs)
```

---

## 1. Evaluation Episodes (`RoboTwin/` and `LIBERO/`)

Each top-level folder groups episodes by **policy (configuration)** and then by
**task**. Every episode records the **first frame of the rollout** (the initial
scene observation) and a JSON file with ground-truth metadata. These episodes
support the *Failure Reason* and *Failure Stage* evaluation tasks described in
the paper.

### 1.1 RoboTwin episodes

`RoboTwin/<policy>/<task_name>/episode_XXX/`

- `first_frame.png` — RGB image of the initial scene (first frame of the rollout).
- `metadata.json` — ground-truth metadata:

```json
{
  "task": "open_laptop",
  "seed": 200000,
  "success": false,
  "instruction": "Pull the lid of the black laptop with smooth matte surface upward.",
  "split": "seen",
  "stages": ["Reach", "Grasp", "Lift"]
}
```

| Field | Description |
|---|---|
| `task` | RoboTwin task name (e.g. `open_laptop`, `place_cans_plasticbox`). |
| `seed` | Random seed used to generate the initial scene. |
| `success` | Ground-truth outcome of the rollout in the simulator (`true`/`false`). |
| `instruction` | Natural-language task instruction given to the policy. |
| `split` | Generalization split of the scene (`seen`). |
| `stages` | Ordered task-specific stage decomposition used for failure-stage annotation (full lists in [Section 3](#3-task-stage-decomposition)). |

- Policies: `pi05`, `pi05-data-20`, `pi05-data-50`, `fastwam`, `lingbot-vla-v2`.
- 20 tasks per policy, 50 episodes per task → 5,000 episodes in total
  (1,774 success / 3,226 failure).

### 1.2 LIBERO episodes

`LIBERO/<policy_variant>/task_XX/episode_XXX/`

- `first_rgb.png` — RGB image of the initial scene (first frame of the rollout).
- `first_rgb.npy` — the same first frame as a raw NumPy array.
- `scene.json` — ground-truth metadata:

```json
{
  "task": "LIVING_ROOM_SCENE2_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket",
  "init_state_id": 0,
  "language": "put both the alphabet soup and the tomato sauce in the basket",
  "suite": "libero_10",
  "success": false,
  "task_id": 0,
  "stages": [
    "Grasp the alphabet soup.",
    "Move the alphabet soup to the basket.",
    "Release the alphabet soup and retreat the arm.",
    "Grasp the tomato sauce.",
    "Move the tomato sauce to the basket.",
    "Release the tomato sauce and retreat the arm."
  ]
}
```

| Field | Description |
|---|---|
| `task` | Full LIBERO task name (scene + instruction). |
| `task_id` | Integer task index within the suite (`task_00`–`task_09`). |
| `suite` | LIBERO suite (`libero_10`, a.k.a. LIBERO-Long). |
| `init_state_id` | Index of the initial state used for the rollout. |
| `language` | Natural-language task instruction. |
| `success` | Ground-truth outcome of the rollout (`true`/`false`). |
| `stages` | Ordered task-specific stage decomposition used for failure-stage annotation (full lists in [Section 3](#3-task-stage-decomposition)). |

- Policy variants: `pi05_data20`, `pi05_data50`, `pi05_data100` (π0.5
  fine-tuned with 20% / 50% / 100% of the training data).
- 10 LIBERO-10 tasks per variant, 50 episodes per task → 1,500 episodes in
  total (1,056 success / 444 failure).

---

## 2. Failure Type Taxonomy

The *Failure Reason* task annotates every confirmed failure with **one primary
failure type**, adapting the planning/execution distinction of
[REFLECT](https://robot-reflect.github.io/) and Guardian to trajectory
annotation. Labels describe *observable behavior* in the context of the task
instruction, without assuming internal policy causes. There are **three
labels**:

| Label | Definition | Examples |
|---|---|---|
| **Task Planning Failure** | Task-inappropriate choices of targets, subgoals, or their ordering — e.g. omitting a required step, or attempting a subgoal before its prerequisites are satisfied. | Closing a cabinet before inserting the required object, then attempting insertion through the closed door (violates a task dependency even if the closing motion itself succeeds). |
| **Action Execution Failure** | Failure to carry out an *intended* subgoal through motion or physical interaction. | Missing a grasp (gripper closes beside the mug handle and lifts without the mug); dropping a held object during transport; inaccurate placement. |
| **Indeterminate Failure** (`Unjudgable`) | A confirmed failure whose primary type cannot be identified reliably, e.g. because the decisive interaction is occluded. Recorded as `Unjudgable`; these episodes remain in the all-episode denominator. | — |

Annotation rules (see Appendix B of the paper):

- For each confirmed failure, annotators assign **one primary type** and **one
  decisive stage**; when several errors occur, the type follows the *earliest
  clearly supported primary cause*, and the stage identifies the *earliest
  decisive event* preventing completion.
- Stage names are copied verbatim from the task-specific stage lists below; if
  a stage cannot be identified, annotators record it as `unknown`.
- Successful episodes contribute zero failure counts.

---

## 3. Task Stage Decomposition

The `stages` field of every episode lists the ordered semantic stages of its
task. These task-specific lists define the **vocabulary of failure stages**:
stage-level failure annotations always reference one of these stages (or
`unknown`). The same decomposition is shared by all policies for a given task.

### 3.1 RoboTwin tasks (20)

Stages of dual-arm tasks are prefixed with the acting arm (`Arm 1:` / `Arm 2:`);
`/` separates stages executed by the two arms in parallel.

| Task | Ordered stages |
|---|---|
| `beat_block_hammer` | Reach → Grasp → Transfer → Strike |
| `blocks_ranking_rgb` | Reach → Grasp red → Transfer right → Place → Grasp green → Transfer center → Grasp blue → Transfer left |
| `blocks_ranking_size` | Reach → Grasp small → Transfer right → Place → Grasp medium → Transfer center → Grasp large → Transfer left |
| `handover_block` | Arm 1: Reach → Arm 1: Grasp → Arm 1: Transfer → Arm 2: Reach → Arm 2: Grasp → Arm 1: Release → Arm 2: Transfer → Arm 2: Place |
| `hanging_mug` | Arm 1: Reach → Arm 1: Grasp → Arm 1: Place → Arm 2: Reach → Arm 2: Grasp → Arm 2: Transfer → Arm 2: Place |
| `move_pillbottle_pad` | Reach → Grasp → Transfer → Place |
| `open_laptop` | Reach → Grasp → Lift |
| `pick_diverse_bottles` | Arm 1: Reach / Arm 2: Reach → Arm 1: Grasp / Arm 2: Grasp → Arm 1: Transfer / Arm 2: Transfer |
| `place_a2b_left` | Reach → Grasp A → Place A |
| `place_bread_skillet` | Arm 1: Reach / Arm 2: Reach → Arm 1: Grasp / Arm 2: Grasp → Rotate the skillet → Transfer the bread → Place the bread |
| `place_cans_plasticbox` | Arm 1: Reach / Arm 2: Reach → Arm 1: Grasp / Arm 2: Grasp → Arm 1: Transfer → Arm 1: Place → Arm 1: Retract → Arm 2: Transfer → Arm 2: Place → Arm 2: Retract |
| `place_dual_shoes` | Arm 1: Reach / Arm 2: Reach → Arm 1: Grasp / Arm 2: Grasp → Arm 1: Transfer → Arm 1: Place → Arm 1: Retract → Arm 2: Transfer → Arm 2: Place |
| `place_empty_cup` | Reach → Grasp → Transfer → Place |
| `place_object_scale` | Reach → Grasp → Transfer → Place |
| `place_phone_stand` | Reach → Grasp → Transfer → Place |
| `put_bottles_dustbin` | Bottle 1: Reach → Bottle 1: Grasp → Bottle 1: Transfer → Bottle 1: Place → Bottle 2: Reach → Bottle 2: Grasp → Bottle 2: Transfer → Bottle 2: Place → Bottle 3: Reach → Bottle 3: Grasp → Bottle 3: Transfer → Bottle 3: Place |
| `put_object_cabinet` | Arm 1: Reach → Arm 1: Grasp → Arm 2: Reach → Arm 2: Grasp → Arm 2: Pull open → Arm 1: Transfer → Arm 1: Place |
| `scan_object` | Arms 1 and 2: Reach → Arms 1 and 2: Grasp → Arm 1: Transfer → Arm 2: Transfer |
| `stack_blocks_three` | Block 1: Reach → Block 1: Grasp → Block 1: Place → Block 1: Retract → Block 2: Reach → Block 2: Grasp → Block 2: Place → Block 2: Retract → Block 3: Reach → Block 3: Grasp → Block 3: Place → Block 3: Retract |
| `stack_bowls_three` | Bowl 1: Reach → Bowl 1: Grasp → Bowl 1: Place → Bowl 1: Retract → Bowl 2: Reach → Bowl 2: Grasp → Bowl 2: Place → Bowl 2: Retract → Bowl 3: Reach → Bowl 3: Grasp → Bowl 3: Place → Bowl 3: Retract |

### 3.2 LIBERO-10 tasks (10)

| Task | Ordered stages |
|---|---|
| task_00 | Grasp the alphabet soup → Move the alphabet soup to the basket → Release the alphabet soup and retreat the arm → Grasp the tomato sauce → Move the tomato sauce to the basket → Release the tomato sauce and retreat the arm |
| task_01 | Grasp the cream cheese box → Move the cream cheese box to the basket → Release the cream cheese box and retreat the arm → Grasp the butter → Move the butter to the basket → Release the butter and retreat the arm |
| task_02 | Turn on the stove → Grasp the moka pot → Move the moka pot to the stove → Place the moka pot on the stove, release it, and retreat the arm |
| task_03 | Open the bottom drawer → Grasp the black bowl → Move the black bowl into the bottom drawer → Release the bowl and retreat the arm → Close the bottom drawer and retreat the arm |
| task_04 | Grasp the white mug → Move the white mug to the left plate → Release the white mug and retreat the arm → Grasp the yellow-and-white mug → Move the yellow-and-white mug to the right plate → Release the mug and retreat the arm |
| task_05 | Grasp the book → Move the book to the back compartment of the caddy → Release the book and retreat the arm |
| task_06 | Grasp the white mug → Move the white mug to the plate → Release the mug and retreat the arm → Grasp the chocolate pudding → Move the chocolate pudding to the right of the plate → Release the pudding and retreat the arm |
| task_07 | Grasp the alphabet soup → Move the alphabet soup to the basket → Release the soup and retreat the arm → Grasp the cream cheese box → Move the cream cheese box to the basket → Release the box and retreat the arm |
| task_08 | Grasp the first moka pot → Move the first moka pot to the stove → Release the first moka pot and retreat the arm → Grasp the second moka pot → Move the second moka pot to the stove → Release the second moka pot and retreat the arm |
| task_09 | Open the microwave door → Grasp the yellow-and-white mug → Move the mug into the microwave → Release the mug and retreat the arm → Close the microwave door and retreat the arm |

---

## 4. Boundary Cases (`boundary/`)

The `boundary/` subset supports the *Failure Boundary* evaluation task. Each
**pair** is constructed by starting from a base initial scene and perturbing
**a single scene dimension** — object `position`, `orientation`, or `size` —
in small steps until the policy's outcome flips. A pair therefore contains two
adjacent variants that straddle the success→failure transition:

- `success/` — rollouts from the last scene variant that (mostly) succeeds;
- `failure/` — rollouts from the first scene variant where performance collapses.

Unlike the evaluation episodes above, every boundary episode also includes the
**full rollout video**, enabling frame-level comparison against world-model
generations.

```
boundary/<Simulator>/<dimension>/pair_XXX/
├── metadata.json                 # pair-level boundary description
├── success/
│   ├── episode_000/              # LIBERO sides: episode_000 - episode_009
│   │   ├── rollout.mp4           # full rollout video
│   │   └── scene.json            # episode-level record
│   └── ...                       # 10 episodes per side
│                                 # (RoboTwin sides are numbered episode_001 - episode_010)
├── failure/
│   └── ...
└── success_scene.json            # (RoboTwin only) full scene snapshot of the success variant
└── failure_scene.json            # (RoboTwin only) full scene snapshot of the failure variant
```

Counts: **RoboTwin** 60 pairs (position / orientation / size × 20) with 10
rollouts per side; **LIBERO** 60 pairs (orientation / position / size × 20)
with 10 rollouts per side → **2,400 boundary
rollout episodes** in total. The pair counts match the paper exactly: 60
frozen boundary pairs per platform, 20 per perturbation factor (Appendix C).
The LIBERO pair list and selection criteria are recorded in
`boundary/LIBERO/manifest.json`.

### 4.1 Episode-level record (`scene.json`)

RoboTwin example:

```json
{
  "policy": "FastWAM",
  "task": "place_cans_plasticbox",
  "language": "Lift the hand-sized metallic can and the white patterned brown metal can, and stick them into the rectangular storage box with grooves",
  "variant_label": "plasticbox_size+0pct",
  "perturb_dimension": "Size",
  "success": false
}
```

LIBERO example:

```json
{
  "policy": "pi05_head_finetuned",
  "task": "KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it",
  "language": "turn on the stove and put the moka pot on it",
  "variant_idx": 2,
  "variant_label": "moka_yaw+20",
  "perturb_dimension": "orientation",
  "perturb_delta": 20.0,
  "success": true,
  "gripper_action_conversion": "unit_to_native"
}
```

| Field | Description |
|---|---|
| `policy` | Policy that was rolled out. |
| `task` / `language` | Task name and natural-language instruction. |
| `variant_label` | Human-readable label of the perturbed scene variant (e.g. `moka_yaw+20`, `plasticbox_size+20pct`). |
| `variant_idx` | (LIBERO) Index of the variant along the perturbation sweep. |
| `perturb_dimension` | Perturbed dimension: `position`, `orientation`, or `size`. |
| `perturb_delta` | (LIBERO) Magnitude of the perturbation relative to the base scene. |
| `success` | Observed outcome of this rollout. |
| `gripper_action_conversion` | (LIBERO) Gripper action space conversion used by the policy. |

### 4.2 Pair-level record (`metadata.json`)

LIBERO example:

```json
{
  "task_id": 2,
  "base_episode": 0,
  "perturb_joint": "moka_pot_1_joint0",
  "perturb_axis": 0,
  "dimension": "position",
  "success_variant": 0,
  "failure_variant": 1,
  "success_delta": 0.0,
  "failure_delta": 0.02,
  "success_label": "moka_x+0cm",
  "failure_label": "moka_x+2cm",
  "success_count": 10,
  "failure_count": 7,
  "success_rate": 1.0,
  "failure_rate": 0.7,
  "asset_id": "libero/pi05_libero_data20"
}
```

| Field | Description |
|---|---|
| `task_id` / `base_episode` | Task index and base episode from which the sweep started. |
| `perturb_joint` / `perturb_axis` | Perturbed object joint and axis. |
| `dimension` | Perturbed dimension (`position` / `orientation` / `size`). |
| `success_variant`, `failure_variant` | Variant indices forming the boundary pair. |
| `success_delta`, `failure_delta` | Perturbation magnitudes of the two variants. |
| `success_label`, `failure_label` | Human-readable labels of the two variants. |
| `success_count`, `failure_count` | Number of successful rollouts observed on each side. |
| `success_rate`, `failure_rate` | Observed success rate on each side of the boundary. |
| `asset_id` | Identifier of the policy variant used for the sweep (e.g. `libero/pi05_libero_data20`). |

**RoboTwin pair records use a different schema** (they do not contain the
LIBERO-specific fields above such as `base_episode`, `perturb_joint`,
`success_variant` or `asset_id`). Example (`RoboTwin/size/pair_041`):

```json
{
  "policy": "FastWAM",
  "policy_id": "fastwam",
  "task_id": 26,
  "task": "place_cans_plasticbox",
  "dimension": "Size",
  "category": "size",
  "object_axis": "plasticbox_size",
  "success_label": "plasticbox_size+0pct",
  "failure_label": "plasticbox_size+20pct",
  "success_rate_high": 0.7,
  "success_rate_low": 0.2,
  "drop": 0.5,
  "endpoint_gap": {
    "previous_value": 1.0,
    "failed_value": 1.2,
    "signed_change": 0.2,
    "absolute_change": 0.2,
    "unit": "ratio",
    "failed_over_previous_size_ratio": 1.2,
    "relative_linear_size_change_percent": 20.0,
    "relative_linear_size_change_absolute_percent": 20.0,
    "scale_factor_reference": "original object linear dimensions"
  }
}
```

| Field | Description |
|---|---|
| `policy`, `policy_id` | Policy that was rolled out (`FastWAM` / `Pi0.5` / `LingBot-VLA-v2`) and its identifier. |
| `task_id`, `task` | RoboTwin task index and name. |
| `dimension` | Perturbed dimension, capitalized: `Position`, `Orientation`, or `Size`. |
| `category` | Lowercase category of the perturbation: `position`, `rotation` (for orientation), or `size`. |
| `object_axis` | Perturbed scene axis (e.g. `plasticbox_size`). |
| `success_label`, `failure_label` | Labels of the two boundary variants. |
| `success_rate_high`, `success_rate_low` | Observed success rates of the success-side and failure-side variants (10 rollouts each). |
| `drop` | Success-rate drop across the boundary (`success_rate_high` − `success_rate_low`). |
| `endpoint_gap` | Physical change between the two variants. Its keys are dimension-specific: position pairs include `absolute_change_cm`, orientation pairs include `absolute_change_degrees`, and size pairs include the size-ratio fields shown above. |

### 4.3 Scene snapshots (RoboTwin only)

`success_scene.json` / `failure_scene.json` store the complete simulator scene
state (object poses, geometry, etc., under `snapshot.objects`) together with
the perturbed `actor`, `feature`, `value`, `unit`, and `instruction`, so that
both boundary scenes can be **exactly re-instantiated** in the RoboTwin
simulator.

---

## 5. Intended Use

The dataset provides the simulator ground truth for the three WorldFailBench
evaluation tasks:

1. **Failure Reason** — given the first frame and instruction of a *failed*
   episode (`RoboTwin/`, `LIBERO/`), evaluate whether a world model's rollout
   reproduces the observable error the real policy makes, using the failure
   type taxonomy in [Section 2](#2-failure-type-taxonomy).
2. **Failure Stage** — using the task-specific stage decompositions
   ([Section 3](#3-task-stage-decomposition)) stored in each episode's
   `stages` field, evaluate whether failures in world-model rollouts
   concentrate at the same execution stages as in the simulator.
3. **Failure Boundary** — using the paired variants in `boundary/`, evaluate
   whether a world model reproduces the success→failure transition under the
   same controlled scene change.

---

## 6. Notes

- All episodes are **closed-loop rollouts of real VLA policies executed in the
  physics simulators**, not in world models; they serve as reference ground
  truth.
- `success` labels are produced by each simulator's task success checker.
- Failure-type and failure-stage labels follow the annotation protocol of the
  paper: stage names are copied from the task-specific lists in
  [Section 3](#3-task-stage-decomposition), and unidentifiable stages are
  recorded as `unknown`; unclassifiable failures are recorded as `Unjudgable`.
- Boundary pairs are labeled by the *relative* outcome change across the pair;
  the absolute success rate on the `failure` side may still be non-zero (see
  `failure_rate` in the pair metadata).

## 7. Citation

If you find this dataset useful, please cite:

```bibtex
@article{worldfailbench,
  title  = {WorldFailBench: Do World Models Understand Policy Failures?},
  author = {Zhang, Zheng and Tang, Xiaoyang and Cheng, Shuaiqi and Huang, Jiakai and Wu, Zhongbo and Zhang, Ziyi and Si, Yuan and Liu, Zeyu and Gao, Weibo and Liu, Shunyu and You, Shan and Wang, Fei and Wang, Xiaogang and Tao, Dacheng},
  year   = {2026}
}
```

## 8. License

This dataset is released under the
[Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).
The full license text is available in [`LICENSE`](LICENSE).