File size: 6,116 Bytes
4c0d9a7 d92b41f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 | ---
license: cc-by-4.0
pretty_name: AGOS-Dataset
tags:
- robotics
- simulation
- vision-language
---
# AGOS-Dataset
AGOS-Dataset contains 2,125 simulated cooperative vehicle-search tasks in CarlaAir and expert observations and actions for the 1,100 training tasks. Each task pairs an aerial robot with a ground vehicle. Both use reference images to identify a vehicle in a simulated town.
## Task sets
| Volume | Tasks | Contents |
| --- | ---: | --- |
| `taskset-train1100` | 1,100 | Training tasks linked to the final expert trajectory census |
| `taskset-validation275` | 275 | Validation tasks in five training towns |
| `taskset-test300` | 300 | Unseen-town test tasks in Town03, with 100 per difficulty |
| `taskset-seentest450` | 450 | Seen-town test tasks, with 150 per difficulty |
The split index lists the episode keys and the SHA-256 digest of the selection manifest for each split. Seen-Test450 uses 450 slots from a certified pool of 523. Its 73 reserve slots are excluded. Train1100 uses the 11 November 2026 audited roster, including its 17 revised source trajectories. Sixteen of these revisions changed the episode key. [Split index](agos-tasksets-splits-v1.json)
A task is identified by its original plan SHA-256. All 2,125 original hashes are distinct. Train1100 and Val275 share 216 map/episode key pairs, each with different plans. All other split pairs have disjoint keys. Join on `source_plan_sha256` or on `(split, episode_key)`. The latter also identifies both reference images. [Task index](tasks.csv)
Each task archive extracts into its own split directory. For example, the first Unseen-Test300 task uses these three files under the data root.
```text
test300/dataset/Town03/lite_easy_00001_plan.json
test300/dataset/Town03/images/lite_easy_00001_target_reference.png
test300/dataset/Town03/images/lite_easy_00001_teammate_reference.png
```
Each split also includes its own `tasks.csv` and `manifest.json`. In the index, `plan_sha256` covers the published JSON bytes. `source_plan_sha256` covers the original plan. Published plans refer to calibration and navigation assets using paths relative to the AGOS-Bench repository root. The `navigation_binding_sha256` and `navigation_certificate_sha256` columns identify the embedded navigation oracle. The publisher recomputed both canonical JSON hashes for every task and checked both reference image hashes against the original plan.
To copy a verified subset or generate a custom ground-start variant on CarlaAir, follow the [task generation guide](https://github.com/Yomier/AGOS-Bench/blob/main/benchmark/MINTING.md). Custom variants have a separate task identity.
## Expert trajectories
The `trajectories-train1100` archive extracts into a directory of the same name. It contains `rows.jsonl`, `episodes.csv`, and `images/<sha256>.png`. The 27,720 rows contain 82,331 image references to 45,843 distinct images. Of these rows, 23,529 belong to the aerial robot and 4,191 to the ground vehicle. The trajectory outcomes were 910 successful confirmations, 177 budget terminations, and 13 navigation stalls.
Each JSONL row has `messages`, `images` and `meta`. Resolve image paths relative to the trajectory directory. `meta.episode_key` and `meta.call_id` identify a call within its episode. When joining to `episodes.csv`, normalize the simulator map spellings `Town04_Opt` and `Town05_Opt` to `Town04` and `Town05`. The CSV links each episode to the training plan by `source_plan_sha256`. The published JSONL preserves the audited row bytes, SHA-256 `24c64ec1b60a74c3d9512772f30b40cae95f0a431c8d38757fe24bbb3a01904c`.
These are simulator expert demonstrations. The historical trained models consumed 6,014 rows after length filtering from an earlier 8,212-row input. Training on the complete 27,720-row release is a distinct training setting.
## Independent certification
The `certification` archive contains six evidence files for each of the 2,125 tasks. Under `certification/<split>/<episode_key>/`, the replay records the independent execution and links to its arrival image. The `oracle_qa/` directory contains the deployed navigation oracle, the deployment bindings before and after execution, and the validation receipt.
The archive's `certification/manifest.json` joins the six files to each published plan and records original and distributed file hashes. For 123 training tasks renamed during dataset construction, `replay_episode_key` and `replay_plan_sha256` identify the plan used in the independent replay. Reapplying the documented rename, reference path changes and oracle rehash reproduced each final source plan byte for byte. Replay records preserve the original engine receipt digest. The arrival image and all four oracle and deployment files preserve their original bytes.
## Access and verification
Download any split on its own, or retrieve all six archives. The four task splits contain 2,125 distinct tasks. The independent certification archive and the training trajectory archive are optional for a standard evaluation. [Release checksums](SHA256SUMS)
```bash
python -m pip install huggingface-hub
hf download Yomier/AGOS-Dataset --repo-type dataset --include 'taskset-test300.tar.gz' 'SHA256SUMS' --local-dir agos-data
cd agos-data
grep ' taskset-test300.tar.gz$' SHA256SUMS | sha256sum --check - && tar -xzf taskset-test300.tar.gz
```
For a complete snapshot, omit `--include` and run `sha256sum --check SHA256SUMS` in the download directory. To reproduce a specific release, add `--revision` and the commit ID shown in the Hub history. [Dataset files](https://huggingface.co/datasets/Yomier/AGOS-Dataset/tree/main)
## License and citation
The dataset is distributed under [Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/). Credit AGOS-Dataset and its authors when redistributing the tasks or demonstrations. The accompanying code has its own license in the AGOS-Bench repository.
Citation metadata is maintained with the code release. The revised manuscript is arXiv:2609.08402v2. [AGOS-Bench](https://github.com/Yomier/AGOS-Bench)
|