JavideuS/Spooky-benchmark

Spooky MAPF Benchmark Sweeps

Aggregated benchmark results for Spooky, a hybrid quantum–classical framework for Multi-Agent Path Finding (MAPF) via QUBO. Each entry here is one sweep — a full instance × problem × solver × ablation matrix run through Spooky's BenchmarkRunner and aggregated into CSV tables.

• Code: https://github.com/JavideuS/Spooky
• Paper: Scalable Multi-Robot Path Planning via Quadratic Unconstrained Binary Optimization, arXiv:2602.14799
• Served by: the Spooky FastAPI Space, whose /v1/analysis/* endpoints snapshot_download this repo (allow_patterns=["sweeps/**", "published.json"]) and render Plotly figures from the CSVs on demand.

Solvers compared across sweeps: ilp (HiGHS), cbs (Conflict-Based Search), dwave (simulated annealing via neal, no QPU), pennylane (simulated QAOA), and, where a sweep enabled hardware, real QPU / IBM backends.

Repository layout

published.json                       ledger — one entry per published sweep
sweeps/
└── <sweep_id>/
    ├── index.json                   the run plan: one entry per matrix cell
    ├── manifest.json                reproducibility record (git, deps, hardware, seed)
    └── analysis/
        ├── runs_long.csv            one row per individual solver run
        ├── summary_by_solver.csv    success rate + variable-reduction means
        ├── statistical_tests.csv    paired Wilcoxon signed-rank tests
        ├── energy_diagnostics.csv   per-config energy-landscape health
        ├── failure_causes.csv       invalid runs split by cause
        ├── robot_statistics_long.csv one row per (run, robot); empty if the
        │                             sweep ran below BenchmarkRunner level 2
        └── benchmark_table.tex      paper-style LaTeX table snippet

Not included

• Raw per-cell benchmark_*.json — they carry full decoded robot paths and run to tens of MB per sweep. The CSVs are sufficient to serve and to reproduce every figure.
• Rendered plots (analysis/plots/) — see Regenerating the figures below. A ~30 MB sweep directory ships here as a few hundred KB.

published.json

Object keyed by sweep_id. Each value:

Field	Meaning
commit, branch	Spooky git state the sweep ran against.
published_at	UTC ISO timestamp of the upload.
start_time, end_time	From the sweep's manifest.json. A null end_time means the sweep was interrupted and its matrix is incomplete.
n_completed	Matrix cells with a completed benchmark JSON.
solvers, problems, grid_sizes	Coverage of the sweep.
dropped_solvers	Solvers filtered out of the CSVs before upload (--drop-solver), or null.

Convention: normally one sweep per Spooky commit. Two different sweep configs on one commit is legitimate and they publish side by side; a re-run of the same config replaces the prior sweep for that commit.

How to read the tables

Full column dictionaries and methodology are in the code repo:

• quantum/benchmark/analysis/README.md — per-column reference for every CSV here.
• quantum/benchmark/README.md — why the paired Wilcoxon signed-rank test, the log transform on time metrics, the rank-biserial effect size, Benjamini–Hochberg correction, and the survivorship/censoring accounting.

Pin those to the sweep's own commit — sweeps/<sweep_id>/manifest.json → git.commit — for the doc version that matches the numbers: https://github.com/JavideuS/Spooky/blob/<commit>/quantum/benchmark/analysis/README.md

Before trusting any row of statistical_tests.csv:

• Use p_value_bh (FDR-adjusted across the whole family of tests), not the raw p_value.
• Read it next to summary_by_solver.csv. A paired test only compares instances both solvers solved, so a solver that only succeeds on easy instances is compared only on easy instances. Check n_dropped_invalid, n_dropped_unshared, min_valid_runs, n_censored and success_rate.
• n_effective (not n_pairs) is the test's real sample size; both must be ≥ 6 or there is no p-value.
• Energy is not in the cross-solver tests — a QUBO Hamiltonian value, an ILP objective and a CBS sum-of-costs are unrelated quantities. Use the within-solver energy_excess column in runs_long.csv for energy.

Regenerating the figures

Only the tables are stored. Every figure is a pure function of runs_long.csv (+ robot_statistics_long.csv). No re-solving is needed.

# 1. Install Spooky's analysis + plotting extras (pin to the sweep's commit)
pip install "spooky[benchmark,visualizer] @ git+https://github.com/JavideuS/Spooky@<commit>"

# 2. Pull one sweep directory from this dataset
hf download JavideuS/Spooky-benchmark --repo-type dataset \
    --include "sweeps/<sweep_id>/**" --local-dir ./spooky-sweeps

# 3. Render the plots into <sweep_dir>/analysis/plots/
python -m quantum.benchmark.analysis.run_plots \
    -d ./spooky-sweeps/sweeps/<sweep_id>

Output (HTML always; PNG/PDF/SVG when kaleido is present — it is in the visualizer extra):

File	Content
scaling.{html,png,pdf,svg}	Execution time vs problem size, per solver.
success_rate.{…}	Valid-solution rate per solver × instance.
variable_reduction.{…}	Pre-processing variable reduction per solver.
energy_excess.{…}	Within-solver energy spread above best (QUBO backends).
path_efficiency.{…}	Per-robot path efficiency — only if robot_statistics_long.csv is non-empty (sweep run at level ≥ 2).

python -m quantum.benchmark.analysis.run_report -d <sweep_dir> also re-derives benchmark_table.tex alongside the plots.

Reproducing a sweep from scratch

sweeps/<sweep_id>/manifest.json embeds the entire sweep config under sweep_config, plus the git commit, package versions, hardware, and global_seed.

git clone https://github.com/JavideuS/Spooky && cd Spooky
git checkout <commit-from-manifest>
pip install -e ".[benchmark,classical,visualizer]"
spooky-sweep --config sweep_configs/<name>.yaml     # same seed → same run
python -m quantum.benchmark.analysis.run_aggregate -d results/sweeps/<new_id>

Classical-only configs (ilp, cbs, simulated dwave/pennylane) need no tokens or quota and are fully reproducible from the seed. QPU/IBM rows are not bit-reproducible and are gated behind explicit flags in spooky-sweep.

License

Apache-2.0, matching the Spooky code repository.

Citation

@misc{gonzalezvillasmil2026scalablemultirobotpathplanning,
      title={Scalable Multi-Robot Path Planning via Quadratic Unconstrained Binary Optimization},
      author={Javier González Villasmil et al.},
      year={2026},
      eprint={2602.14799},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2602.14799},
}