---
license: other
task_categories:
  - robotics
  - other
language:
  - en
tags:
  - embodied-ai
  - tracking
  - dagger
  - habitat
pretty_name: OpenTrackVLA DAgger Recovery Dataset
---

# OpenTrackVLA DAgger Recovery Dataset

Training-only DAgger recovery data generated in Habitat for OpenTrackVLA target
tracking. This release contains V6-policy failure recovery data for AT, DT, and
STT, using seeds 101 and 102.

The uploaded payload contains only the final training-ready datasets:

- JSONL trajectory labels;
- referenced RGB frames;
- precomputed DINOv3 + SigLIP fine/coarse visual-token caches;
- conversion, filtering, and integrity metadata.

Source rollouts, raw recovery videos, evaluation-scene data, model checkpoints,
and EVT-Bench results are not included.

## Layout

```text
packed/v6_failures/train/<task>/seed_<seed>/
  jsonl.tar.zst.part-0000...
  frames.tar.zst.part-0000...
  vision_cache.tar.zst.part-0000...
metadata/v6_failures/train/<task>/seed_<seed>/
  dataset_stats.json
  dataset_inspection.json
  dagger_selection_plan.json
  referenced_frames.txt
  *.manifest.json
```

Each component is a Zstandard-compressed tar stream split into numbered parts.
Reconstruct one component with:

```bash
cat <component>.tar.zst.part-* | tar --zstd -xf - -C /path/to/dataset
```

Every component manifest records the ordered part names, byte sizes, and SHA256
checksums. Verify all parts before extraction.

## Data policy

This dataset is derived from simulated Habitat/HM3D tracking episodes. Users are
responsible for complying with the licenses and terms of the upstream scene,
avatar, simulator, and model assets. The `other` license marker is intentional:
it does not replace those upstream terms.

## Generation

- Split: training scenes and training avatars only
- Tasks: AT, DT, STT
- Seeds: 101, 102
- History: 31 frames
- Prediction horizon: 8 waypoints
- DAgger target fraction: 0.25
- Failure window: 12 steps
- Intervention pre/post windows: 12/8 steps
- Maximum retained samples per episode: 26

Deterministic Habitat native-process crashes are explicitly audited and excluded;
they are never converted into synthetic success/failure labels.