# FailBench

From the catalogue. Listed by one researcher from one source page.

- id: `failbench`
- kind: language and reasoning annotation
- robot or device: 14 public sources, 12 real-world and 2 simulated (class: cross embodiment)
- how collected: curated policy attempts with failure labels
- size as stated: 2,197 manipulation attempts
- hours: not stated | hours only claimed: none | episodes: not stated
- year: 2026
- organisation: not stated
- licence: not stated (class: not stated)
- access: not stated (class: not stated)
- link (paper): https://arxiv.org/abs/2609.03611
- note: 75 percent of failures occur naturally.

## Copies and ports on the Hugging Face Hub

Uploads whose name carries this dataset's name, most downloaded first. Each has its own licence, which can differ from the original's.

| repository | robot | class | kind | episodes | hours | licence | downloads (30 d) |
|---|---|---|---|---|---|---|---|
| [aaronngx/failbench-libero-v2](https://datasets.gurasees.com/datasets/aaronngx/failbench-libero-v2.md) | Franka | single arm | other |  |  | open | 572 |
| [aaronngx/failbench-robocasa-v2](https://datasets.gurasees.com/datasets/aaronngx/failbench-robocasa-v2.md) | Franka | single arm | other |  |  | open | 302 |
| [aaronngx/failbench-wm1](https://datasets.gurasees.com/datasets/aaronngx/failbench-wm1.md) | Franka | single arm | recording | 66,480 | 102 | open | 100 |

Confirm the licence at the link before relying on it.