README.md

---
license: cc-by-4.0
task_categories:
  - image-segmentation
  - object-detection
  - video-classification
language: []
tags:
  - robotics
  - tracking
  - articulated-objects
  - point-tracking
  - long-horizon
  - sapien
  - partnet-mobility
  - rgb-d
  - manipulation
  - affordance
  - semantic-drift
  - embodied-ai
  - video
  - depth
  - multi-view
pretty_name: QueST PartNet-Mobility SAPIEN Dataset
size_categories:
  - 10K<n<100K
dataset_info:
  features:
    - name: video
      dtype: video
    - name: affordance_visualization
      dtype: image
    - name: manipulation_level
      dtype:
        class_label:
          names:
            - manipulation_1
            - manipulation_2
            - manipulation_3
            - manipulation_4
    - name: take_id
      dtype: string
    - name: object_id
      dtype: string
    - name: n_joints
      dtype: int32
    - name: split
      dtype: string
    - name: has_depth
      dtype: bool
  splits:
    - name: train
      num_examples: 16000
    - name: test
      num_examples: 2442
---

# QueST: PartNet-Mobility SAPIEN Simulation Dataset

[![Paper](https://img.shields.io/badge/Paper-CAO%40ICLR%202026-blue)](https://arxiv.org/abs/2605.09513v1)
[![License](https://img.shields.io/badge/License-CC--BY%204.0-green)](https://creativecommons.org/licenses/by/4.0/)
[![Downloads](https://img.shields.io/badge/Downloads-2.5K%2Fmonth-brightgreen)]()
[![IIITA](https://img.shields.io/badge/IIIT-Allahabad-orange)](https://www.iiita.ac.in)

This dataset accompanies the paper:

> **QueST: Persistent Queries as Semantic Monitors for Drift Suppression in Long-Horizon Tracking**  
> Mayank Anand, Mohammad Saqlain, Kyan Mahajan, Priya Shukla, G.C Nandi, Andrew Melnik  
> *CAO Workshop at ICLR 2026*

---

## What Is This Dataset?

Synchronized RGB-D simulation sequences rendered in **SAPIEN** from **PartNet-Mobility** articulated objects, designed to stress-test long-horizon point tracking under articulation, occlusion, and viewpoint change.

The dataset supports the **QueST framework** — which replaces frame-to-frame Markovian tracking with persistent semantic queries that attend globally across time, achieving a **67.7% APE reduction** over TAP-Net on long-horizon articulated sequences.

---

## Exact Folder Structure

Each sequence is stored as an individual `take` folder:

```
QueST-PartNetMobility-SAPIEN/
│
├── manipulation_1/              Level 1 — 1 joint actuated
│   ├── {object_id}/
│   │   ├── take_00/
│   │   │   ├── frames/          RGB-D frames (PNG sequence)
│   │   │   ├── affordance/      Pixel-level affordance maps
│   │   │   ├── video.mp4        Full sequence video (36.6 kB avg)
│   │   │   ├── affordance_vis_10frames.png   Visualization (455 kB)
│   │   │   └── metadata.json    Sequence metadata (20.5 kB)
│   │   ├── take_01/
│   │   └── ...
│   └── ...
│
├── manipulation_2/              Level 2 — 2 joints actuated
├── manipulation_3/              Level 3 — 3 joints actuated
└── manipulation_4/              Level 4 — 4 joints, 240 frames
```

### What Each File Contains

| File | Description |
|---|---|
| `frames/` | Individual RGB-D frames as PNG — use for frame-level tracking evaluation |
| `affordance/` | Pixel-level affordance annotations — interaction-relevant regions labeled |
| `video.mp4` | Full sequence as compressed video — use for temporal model training |
| `affordance_vis_10frames.png` | Visual summary of affordance labels across 10 evenly-spaced frames |
| `metadata.json` | Object ID, joint configuration, ground truth 3D trajectories, camera intrinsics |

---

## Complexity Levels

| Level | Folder | Joints actuated | Max frames | Purpose |
|---|---|---|---|---|
| 1 | `manipulation_1` | 1 | ~60 | Short-horizon training baseline |
| 2 | `manipulation_2` | 2 | ~120 | Medium complexity |
| 3 | `manipulation_3` | 3 | ~180 | Hard multi-joint sequences |
| 4 | `manipulation_4` | **4** | **240** | Long-horizon stress test |

Each level actuates joints **sequentially** — Level 4 is the cumulative long-horizon challenge designed to expose drift in Markovian trackers.

---

## Key Statistics

| Property | Value |
|---|---|
| Total images / rows | 18,442 |
| Total size | 2.27 GB |
| Camera viewpoints | 3 synchronized RGB-D views per sequence |
| Renderer | SAPIEN (physics-based) |
| Depth data | Included in frames/ |
| Annotations | Pixel-level affordance + 3D ground-truth trajectories |
| Articulation types | Revolute, prismatic |
| Object categories | Storage furniture, appliances, hinged devices |

---

## Loading the Dataset

```python
from datasets import load_dataset

# Load full dataset
ds = load_dataset("AnandMayank/QueST-PartNetMobility-SAPIEN")

# Load only long-horizon sequences (manipulation_4)
ds = load_dataset(
    "AnandMayank/QueST-PartNetMobility-SAPIEN",
    data_files={"train": "manipulation_4/**/*"}
)
```

### Loading metadata for a specific take

```python
import json
from huggingface_hub import hf_hub_download

# Download metadata for a specific take
meta_path = hf_hub_download(
    repo_id="AnandMayank/QueST-PartNetMobility-SAPIEN",
    filename="manipulation_1/35059/take_00/metadata.json",
    repo_type="dataset"
)

with open(meta_path) as f:
    meta = json.load(f)

print(meta.keys())
# dict_keys(['object_id', 'joint_config', 'trajectory_gt',
#            'camera_intrinsics', 'affordance_labels', ...])
```

### Loading frames for tracking evaluation

```python
from huggingface_hub import snapshot_download
import os
from PIL import Image

# Download a single take
path = snapshot_download(
    repo_id="AnandMayank/QueST-PartNetMobility-SAPIEN",
    repo_type="dataset",
    allow_patterns="manipulation_4/*/take_00/**"
)

# Load frames in order
frames_dir = os.path.join(path, "manipulation_4/35059/take_00/frames")
frames = sorted([
    Image.open(os.path.join(frames_dir, f))
    for f in os.listdir(frames_dir)
    if f.endswith(".png")
])
print(f"Loaded {len(frames)} frames")
```

---

## Benchmark Results

| Method | APE ↓ | Drift@100 ↓ | Identity Acc ↑ |
|---|---|---|---|
| RAFT-3D | 0.341 | 0.472 | 8.7% |
| CoTracker | 0.276 | 0.398 | 19.2% |
| TAP-Net | 0.251 | 0.372 | 21.4% |
| **QueST (ours)** | **0.081** | **0.155** | **86.5%** |

QueST achieves **67.7% APE reduction** over TAP-Net — the strongest prior method — while maintaining bounded error growth vs near-linear drift in baselines.

---

## Reproducing Results

```bash
git clone https://github.com/AnandMayank/QueST
cd QueST
pip install -r requirements.txt

# Download dataset
python scripts/download_dataset.py \
  --repo AnandMayank/QueST-PartNetMobility-SAPIEN \
  --output data/

# Evaluate on Level 4 long-horizon sequences
python evaluate.py \
  --data data/manipulation_4 \
  --checkpoint checkpoints/quest_full.ckpt \
  --level 4
```

---

## Citation

If you use this dataset please cite:

```bibtex
@inproceedings{anand2026quest,
  title     = {QueST: Persistent Queries as Semantic Monitors for 
               Drift Suppression in Long-Horizon Tracking},
  author    = {Anand, Mayank and Saqlain, Mohammad and Mahajan, Kyan 
               and Shukla, Priya and Nandi, G.C. and Melnik, Andrew},
  booktitle = {CAO Workshop at ICLR 2026},
  year      = {2026}
}
```

---

## License

[Creative Commons Attribution 4.0 International (CC-BY 4.0)](https://creativecommons.org/licenses/by/4.0/)

Free to use for any purpose including commercial use, with attribution.

---

## Contact

**IIIT Allahabad** — Department of Information Technology  
Mayank Anand · [iit2024036@iiita.ac.in](mailto:iit2024036@iiita.ac.in)  
G.C. Nandi · [gcnandi@iiita.ac.in](mailto:gcnandi@iiita.ac.in)

**University of Bremen**  
Andrew Melnik · [andrew.melnik.papers@gmail.com](mailto:andrew.melnik.papers@gmail.com)

Issues and questions: [github.com/AnandMayank/QueST](https://github.com/AnandMayank/QueST)