---
dataset_info:
  config_name: robometer_failure_detection
  features:
  - name: id
    dtype: string
  - name: task
    dtype: string
  - name: lang_vector
    list: float32
  - name: data_source
    dtype: string
  - name: frames
    dtype: string
  - name: is_robot
    dtype: bool
  - name: quality_label
    dtype: string
  - name: preference_group_id
    dtype: 'null'
  - name: preference_rank
    dtype: 'null'
  - name: partial_success
    dtype: 'null'
  splits:
  - name: train
    num_bytes: 175273
    num_examples: 100
  download_size: 28483
  dataset_size: 175273
configs:
- config_name: robometer_failure_detection
  data_files:
  - split: train
    path: robometer_failure_detection/train-*
---

# Robometer Failure Detection Dataset

The exact **100 videos** that the reported Robometer failure-detection results are
computed on: 81 trajectories across 7 tasks, external and wrist views (43 ext / 57
wrist), labels `successful` / `suboptimal` / `failure` (suboptimal counted as failure
at evaluation time).

Tasks: foldtowel, movebanana, movemouse, pourpebble, pulltissue, putspoon, stirpot.

## Metric note

Reported F1 uses **failure as the positive class**, and the "Average" row is pooled over
all videos (not a macro-average over tasks).

## Usage

```python
from datasets import load_dataset

ds = load_dataset("myconnects/robometer_failure_detection",
                  "robometer_failure_detection", split="train")
```