license: apache-2.0
task_categories:
- robotics
- tabular-classification
tags:
- robotics
- failure-detection
- robot-training-data
- annotation-quality
- manipulation
pretty_name: Haptal Robotics Failure Benchmark v1.0
size_categories:
- 1K<n<10K
Haptal Robotics Failure Benchmark v1.0
The first public benchmark for robot training data annotation quality and failure detection in manipulation episodes.
What this is
A held-out test set of 600 robot episodes across 6 failure classes generated from real LeRobot trajectories with physics-based failure injection. The test set is fixed. Anyone can evaluate their annotation pipeline against it and get a comparable score.
Why it exists
No standardized benchmark currently exists for robot training data annotation quality. Open X-Embodiment, BridgeData V2, DROID and LeRobot are dataset collections. None measure annotation quality or failure detection accuracy. We built this to fill that gap.
Failure classes
| Class | Description |
|---|---|
grasp_slip |
Grip force drops causing object slip |
nominal |
Successful episode, no failure |
overcorrect |
Post-failure panic response |
stuck_joint |
Motor stall or joint lock |
trajectory_deviation |
Drift from intended path |
velocity_spike |
Sudden joint velocity anomaly |
Dataset splits
| Split | Episodes |
|---|---|
| Train | 2,400 |
| Test (fixed) | 600 |
Leaderboard
| Rank | Model | Accuracy | Macro F1 | Cohen's κ | OOD F1 | Gap |
|---|---|---|---|---|---|---|
| 🥇 | Haptal (multi-dataset RF) | 93.6% | 0.937 | 0.923 | 0.907 | 0.030 |
| — | Human operator (pass/fail only)* | 83.1% | — | 0.661 | — | — |
| — | Majority baseline | 53.1% | — | 0.000 | — | — |
* Human operators provide binary pass/fail only — no failure type, no timestep.
Submit your model → aarav@haptal.ai · see leaderboard.json for full entry details.
Haptal baseline results
| Metric | Value |
|---|---|
| In-distribution accuracy | 93.6% |
| OOD accuracy (unseen robot) | 90.8% |
| Generalization gap | 0.03 |
| Macro F1 | 0.937 |
| Cohen's Kappa | 0.923 |
| Weakest class F1 | 0.887 (grasp_slip) |
How to use
from datasets import load_dataset
dataset = load_dataset("haptal-ai/robotics-failure-benchmark")
train = dataset["train"]
test = dataset["test"]
How to score your model
- Run your model on the test split
- Save predictions as a CSV with columns
episode_idandpredicted_class - Run
python score.py your_predictions.csv - Email results to aarav@haptal.ai to be added to the leaderboard
Base datasets
Generated from real robot trajectories across multiple platforms:
lerobot/pushtlerobot/xarm_lift_medium_replaylerobot/xarm_push_medium_replaylerobot/aloha_sim_transfer_cube_human
License
Apache 2.0. Base trajectories from LeRobot (MIT license).
Citation
@dataset{haptal2026rfb,
title = {Haptal Robotics Failure Benchmark v1.0},
author = {Bedi, Aarav},
year = {2026},
publisher = {HuggingFace},
url = {https://huggingface.co/datasets/haptal-ai/robotics-failure-benchmark}
}