---
license: apache-2.0
task_categories:
- robotics
tags:
- robotics
- lerobot
- so100
- manipulation
- curated
  - imitation-learning
paper: arxiv:2606.05588
configs:
- config_name: default
  data_files:
  - split: train
    path: episodes.parquet
---

# HaptalAI / SO-100 Curated Manipulation Pack

> A quality-filtered, failure-labelled merge of SO-100 community manipulation datasets.
> Built for imitation learning, RL from demonstrations, and failure-aware training.

## What This Is

The SO-100 open-source robot arm has generated dozens of small community datasets on HuggingFace, each contributed by independent researchers and hobbyists. Individually these datasets are too small and inconsistent to train on reliably. This pack:

- **Merges 13 community SO-100/SO-101 datasets** into a single parquet file
- **Classifies every episode** as `CLEAN` (47) or `FAILURE` (48);   removes `CORRUPTED` episodes (1) entirely
- **Keeps failure episodes** — labelled but not discarded — because failure data   is valuable for reinforcement learning and anomaly-aware imitation
- Provides **full provenance** via `source_dataset` and `original_episode_id` columns

## Curation Methodology

### Why We Keep Failure Episodes

Most curated datasets silently discard failures. We believe this is a mistake:

- **RL from demonstrations** needs to distinguish successful from failed rollouts
- **Failure-conditioned imitation** (DAgger, HG-DAgger) requires negative examples
- **Anomaly detection** for safety systems needs real failure signatures
- **Contrastive training** — failure episodes teach the policy what *not* to do

Failures are flagged with `use_for_training=false` and a `failure_type` tag so downstream users can include or exclude them with a single filter.

### Classification Rules

All classification streams parquet data only — no video is downloaded.

| Label | Condition | Action |
|-------|-----------|--------|
| CORRUPTED | NaN/Inf >50% frames; all channels frozen (std=0); empty episode | **Removed entirely** |
| FAILURE | Velocity spike (z-score > 6.5); action-state mismatch (>50% frames off by >20% of range) | **Kept, `use_for_training=false`** |
| CLEAN | Passes all checks | **Kept, `use_for_training=true`** |

Episode boundaries inferred from `episode_index` column; fallback to fixed 200-frame windows.

## Column Schema

| Column | Type | Description |
|--------|------|-------------|
| `source_dataset` | `string` | HuggingFace dataset ID this episode came from |
| `original_episode_id` | `string` | Episode key as it appeared in the source dataset |
| `frame_index` | `int32` | 0-based timestep index within the episode |
| `use_for_training` | `bool` | `true` = clean; `false` = failure (keep for RL) |
| `failure_type` | `string` | `none` \| `velocity_spike` \| `action_state_mismatch` |
| `quality_score` | `float32` | Quality score \[0,1\]; 1.0 = perfect |
| `episode_length` | `int32` | Total frames in this episode (same for all rows of an episode) |
| `state_0` … `state_6` | `float64` | Joint position / observation state (7 dims for SO-100) |
| `action_0` … `action_6` | `float64` | Commanded action targets (7 dims for SO-100) |

> **Layout**: one row per timestep. Episode metadata (`use_for_training`, `failure_type`, etc.)
> is repeated on every row, enabling direct pandas filtering without joins.

## How to Load

```python
from datasets import load_dataset
import pandas as pd, numpy as np

# Load everything (one row = one timestep)
ds = load_dataset("HaptalAI/so100-curated", split="train")
df = ds.to_pandas()

# Filter: only clean timesteps for BC / imitation learning
clean_df = df[df["use_for_training"] == True]

# Filter: only failure timesteps for RL / anomaly work
fail_df = df[df["use_for_training"] == False]

# Reconstruct one full episode as a trajectory tensor
ep_df = df[(df["source_dataset"] == "cadene/so100_test") &
           (df["original_episode_id"] == "ep_0000")]
state  = ep_df[[f'state_{i}'  for i in range(7)]].to_numpy()   # (T, 7)
action = ep_df[[f'action_{i}' for i in range(7)]].to_numpy()   # (T, 7)

# Or load directly from parquet
df = pd.read_parquet("episodes.parquet")
```

## Stats

| Metric | Value |
|--------|-------|
| Source datasets | 13 |
| Total episodes kept | 95 |
| Clean episodes (`use_for_training=true`) | 47 |
| Failure episodes (`use_for_training=false`) | 48 |
| Corrupted episodes removed | 1 |
| Total timestep rows in parquet | 49,073 |
| Parquet file size | 0.8 MB |
| Generated | 2026-05-24 |

## Attribution

See [sources.md](sources.md) for the full per-dataset breakdown with episode counts, authors, and licenses.

All original data remains under its original license. This curated pack adds quality labels and is released under Apache 2.0.

## Limitations

- Episode classification is heuristic — some edge cases may be mislabelled
- Trajectory columns vary by source; not all episodes have all columns
- Video observations are not included (parquet-only streaming)
- Gated / authentication-required datasets are excluded
- This is a community effort; verify episodes before production training

## Contact

Questions, corrections, or additional source datasets: [aarav@haptal.ai](mailto:aarav@haptal.ai)

Built with ❤️ for the open-source robotics community.