AstroReason-Bench Datasets

This is the canonical Hugging Face Dataset repository for AstroReason-Bench, a benchmark suite for evaluating AI agents and algorithms on space mission design and planning problems.

Each benchmark is exposed as a separate config (subset) within this dataset. Splits within a config map transparently to the benchmark's own dataset splits (e.g., test, single_orbit, multi_orbit).

Dataset Summary

Config	Cases	Splits	Domain
aeossp_standard	5	test	Agile Earth-observation satellite scheduling
regional_coverage	5	test	SAR-like regional strip-observation planning
relay_constellation	5	test	Relay satellite constellation augmentation
revisit_constellation	5	test	Constellation design for uniform target revisit
satnet	5	test	Deep Space Network (DSN) antenna scheduling
spot5	21	single_orbit, multi_orbit, test	SPOT-5 daily photograph scheduling (DCKP)
stereo_imaging	5	test	Optical stereo/tri-stereo imaging planning

Dataset Structure

Every example in every config follows the same schema:

{
  "case_id": "case_0001",
  "split": "test",
  "benchmark": "aeossp_standard",
  "index_metadata": { ... case-specific metadata from index.json ... },
  "files": [
    {"path": "mission.yaml", "content": "..."},
    {"path": "satellites.yaml", "content": "..."},
    {"path": "tasks.yaml", "content": "..."}
  ]
}

• case_id: Unique identifier for the case within the benchmark.
• split: The dataset split the case belongs to.
• benchmark: The benchmark name.
• index_metadata: The case-level entry from the benchmark's dataset/index.json (e.g., satellite counts, task counts, horizons, thresholds, provenance).
• files: A list of all text files inside the case directory. Each entry has a path (relative to the case directory) and the full UTF-8 content.

Note: Because different cases contain different filenames, files is stored as a uniform list of objects rather than a dictionary with dynamic keys. This ensures consistent features across splits.

Quickstart

Loading a single benchmark

from datasets import load_dataset

# Load the aeossp_standard benchmark
ds = load_dataset("AstroReason-Bench/datasets", "aeossp_standard")
print(ds["test"][0]["case_id"])

Loading a specific case's files

case = ds["test"][0]
for file in case["files"]:
    print(file["path"])
    # file["content"] contains the full text of the file

Iterating all configs

from datasets import get_dataset_config_names

configs = get_dataset_config_names("AstroReason-Bench/datasets")
for config in configs:
    ds = load_dataset("AstroReason-Bench/datasets", config)
    for split_name, split_ds in ds.items():
        print(f"{config}/{split_name}: {len(split_ds)} cases")

Benchmark Descriptions

aeossp_standard

A planning-oriented agile Earth-observation satellite scheduling benchmark. Each case provides a fixed constellation of real satellites (via frozen TLEs), time-windowed point-imaging tasks, and hard constraints on observation geometry, battery state, and slew feasibility. The solver submits a schedule of observation actions. Metrics include completion ratio (CR), weighted completion ratio (WCR), time-averaged tardiness (TAT), and power consumption (PC).

Case files: mission.yaml, satellites.yaml, tasks.yaml

regional_coverage

A SAR-like regional imaging benchmark. The solver must plan strip observations over polygonal regions of interest to maximize unique weighted coverage. Cases include real satellites with frozen TLEs, GeoJSON region definitions, and a benchmark-owned fine-grid scoring model. Hard constraints include roll-only strip geometry, same-satellite retargeting limits, battery feasibility, and optional per-region minimum coverage thresholds.

Case files: manifest.json, satellites.yaml, regions.geojson, coverage_grid.json

relay_constellation

A partial constellation-design benchmark for relay service augmentation. Given an immutable MEO relay backbone and ground endpoints, the solver adds a bounded number of lower-altitude relay satellites and schedules ground-link and inter-satellite-link actions. The verifier scores service fraction, latency percentiles, and the number of added satellites.

Case files: manifest.json, network.json, demands.json

revisit_constellation

A constellation-design and scheduling benchmark focused on revisit performance. The solver designs a satellite constellation (initial GCRF Cartesian states up to a case cap) and schedules observation actions to keep target revisit gaps as small as possible over a 48-hour horizon. Scoring is driven by mean_revisit_gap_hours, max_revisit_gap_hours, and satellite_count.

Case files: assets.json, mission.json

satnet

A reinforcement-learning benchmark derived from NASA/JPL Deep Space Network (DSN) operations. The task is to schedule ground-station antenna tracks for interplanetary spacecraft over one-week windows, respecting precomputed view periods, setup/teardown times, maintenance windows, and non-overlap constraints. The primary metric is total scheduled communication hours.

Case files: problem.json, maintenance.csv, metadata.json

spot5

A constraint optimization benchmark based on the ROADEF 2003 Challenge and CNES SPOT-5 operations. Cases are encoded in the DCKP (Disjunctively Constrained Knapsack Problem) format. The solver selects photographs and assigns cameras to maximize total profit while respecting binary/ternary disjunctive constraints and an onboard memory capacity constraint (for multi-orbit instances).

Case files: <case_id>.spot

stereo_imaging

An optical satellite stereo imaging benchmark. The solver schedules timed observations from real satellites to acquire same-pass stereo or tri-stereo products over ground targets. The verifier scores coverage_ratio (fraction of targets with a valid stereo product) and normalized_quality (mean best-per-target quality based on convergence angle, overlap, and pixel scale).

Case files: satellites.yaml, targets.yaml, mission.yaml

Data Splits and Splits Policy

• aeossp_standard, regional_coverage, relay_constellation, revisit_constellation, satnet, stereo_imaging: Currently expose a single committed split test.
• spot5: Exposes three splits:
  • single_orbit: 14 cases without memory constraints.
  • multi_orbit: 7 cases with a memory capacity of 200.
  • test: A 5-case sample drawn with seed 42 (overlaps with single_orbit and multi_orbit).

Future benchmark releases may add additional splits (e.g., train, val) transparently without changing the schema.

Dataset Creation

All canonical datasets are generated or curated by the AstroReason-Bench repository. Where generators exist, they are deterministic and tied to committed splits.yaml contracts. Canonical cases are committed to the repository and are the source of truth for evaluation.

Source Data

Config	Primary Sources
aeossp_standard	CelesTrak TLE snapshot; GeoNames cities; Natural Earth land polygons
regional_coverage	CelesTrak TLE snapshot; GeoNames; Natural Earth
relay_constellation	Synthetic case generator with deterministic seeds
revisit_constellation	Kaggle world-cities dataset; CelesTrak TLE snapshot
satnet	Derived from NASA/JPL Deep Space Network operations research (Chien et al., 2021)
spot5	Mendeley Data DCKP abstraction (Wei & Hao, 2021) of CNES SPOT-5 ROADEF 2003 instances
stereo_imaging	Kaggle world-cities; CelesTrak TLE snapshot

Considerations for Using the Data

• Algorithm-agnostic: Benchmarks define problems and verification, not preferred solving strategies.
• Standalone: Each config is self-contained with no runtime dependencies on other configs.
• No solutions included: This dataset contains only problem instances (cases). Solutions, baselines, and leaderboards belong in downstream repositories.
• Binary files skipped: The upload script ingests only text-based case files. Any future binary artifacts would be excluded from this HF release.

Licensing Information

This dataset repository aggregates multiple sources with different provenance:

• spot5: The .spot instances are from the Mendeley Data release (DOI: 10.17632/2kbzg9nw3b.1) and are provided under CC BY 4.0.
• satnet: Derived from NASA/JPL Deep Space Network operations research. Used for research and educational purposes.
• All other benchmarks (aeossp_standard, regional_coverage, relay_constellation, revisit_constellation, stereo_imaging): Original benchmark materials created by the AstroReason-Bench project.

Please cite the appropriate references when using individual benchmarks (see Citation Information).

Citation Information

If you use this dataset suite in your research, please cite the AstroReason-Bench paper and the original benchmark sources:

AstroReason-Bench (suite)

@article{wang2026astroreason,
  title={AstroReason-Bench: Evaluating Unified Agentic Planning across Heterogeneous Space Planning Problems},
  author={Wang, Weiyi and Chen, Xinchi and Gong, Jingjing and Huang, Xuanjing and Qiu, Xipeng},
  journal={arXiv preprint arXiv:2601.11354},
  year={2026}
}

SatNet

@inproceedings{goh2021satnet,
  title={SatNet: A benchmark for satellite scheduling optimization},
  author={Goh, Edwin and Venkataram, Hamsa Shwetha and Balaji, Bharathan and Wilson, Brian D and Johnston, Mark D},
  booktitle={AAAI-22 workshop on Machine Learning for Operations Research (ML4OR)},
  year={2021}
}

SPOT-5 / DCKP

@article{wei2023responsive,
  title={Responsive strategic oscillation for solving the disjunctively constrained knapsack problem},
  author={Wei, Zequn and Hao, Jin-Kao and Ren, Jintong and Glover, Fred},
  journal={European Journal of Operational Research},
  volume={309},
  number={3},
  pages={993--1009},
  year={2023},
  publisher={Elsevier}
}

Contact and Links

• Repository: https://github.com/Mtrya/AstroReason-Bench
• Issue Tracker: https://github.com/Mtrya/AstroReason-Bench/issues