EcoDrive-FaultBench (dataset)

Synthetic heavy-truck routes, splits, frozen rollout results, timing traces, and Croissant metadata for the EcoDrive-FaultBench benchmark.

This repository hosts the public data bundle for NeurIPS Evaluations & Datasets. Full simulator code, controllers, and experiment drivers live in the separate anonymous code artifact referenced by the paper.

Loading the data

The three result tables ship as Parquet with explicit configurations, so load_dataset returns typed rows:

from datasets import load_dataset

balanced = load_dataset("ecodrive-faultbench-anon/ecodrive-faultbench", "balanced_v2_rollouts")["train"]
delayed  = load_dataset("ecodrive-faultbench-anon/ecodrive-faultbench", "delayed_controller_rollouts")["train"]
routes   = load_dataset("ecodrive-faultbench-anon/ecodrive-faultbench", "route_index")["train"]

len(balanced), len(delayed), len(routes)   # 22750, 5200, 3700
balanced[0]["controller"], balanced[0]["fault_profile"]

The CSV copies remain available for direct download and carry identical content:

import pandas as pd
base = "https://huggingface.co/datasets/ecodrive-faultbench-anon/ecodrive-faultbench/resolve/main/data"
pd.read_csv(f"{base}/balanced_v2_rollouts.csv")

What is included

Path	Rows	Description
data/route_index.csv / .parquet	3,700	Per-route metadata. 1,000 train + 500 validation + 2,200 benchmark.
data/balanced_v2_rollouts.csv / .parquet	22,750	Main protocol: 5 controllers x 7 fault profiles x 650 benchmark routes.
data/delayed_controller_rollouts.csv / .parquet	5,200	Delayed-controller protocol: 2 controllers x 4 fault profiles x 650 benchmark routes.
data/stale_execution_paths.csv	43	Derived labelling of which controller/fault groups share the held-command execution path.
data/external_route_envelope.csv	13	Synthetic route-class envelopes against sampled external reference routes.
data/public_latency_anchors.csv	12	Published latency anchors with measured quantity and mapping into action age.
data/routes.tar.gz	—	routes/ (JSON per route) and metadata/.
data/splits.tar.gz	—	Route IDs per benchmark split.
data/results_frozen_balanced_v2.tar.gz	—	Per-split CSVs under final_v2_postfix/.
data/results_frozen_delayed_controllers.tar.gz	—	Per-split delayed-controller CSVs.
data/traces_sample.tar.gz	—	Scheduler and HIL timing traces for replay experiments.

Documentation lives in dataset_card.md, benchmark_card.md, and model_cards/README.md. Integrity digests are in checksums/SHA256SUMS.txt.

The two result tables answer different questions

balanced_v2_rollouts measures deployed stacks end to end. delayed_controller_rollouts isolates the delivery pipeline: it runs classical PID controllers through the same remote delivery path, so any degradation cannot be attributed to a learned policy.

stale_execution_paths.csv makes the coupling explicit. When a command misses the 100 ms deadline, the deployment stack applies the previously held command instead of the controller's output, so every remotely delivered controller executes the same trajectory regardless of its policy. Under poor_lte, RemoteAman, DelayedPID, and DelayedLookAheadPID therefore agree to six decimal places (71.678101 kWh/km, 84.908058 hard violations/km, pooled over 650 routes), while under ideal the delayed variants reduce exactly to their onboard counterparts. Read held_command_fraction before comparing controllers: where it approaches 1, the comparison measures the delivery path, not the policy.

Aggregate numbers in the paper are macro averages over the seven route classes, not pooled means over runs, because the classes are deliberately unbalanced (50 routes for adversarial_steep, 100 for each other class).

Quick integrity check

sha256sum -c checksums/SHA256SUMS.txt

Croissant

croissant_metadata.json conforms to Croissant 1.1 with the canonical MLCommons @context, declares CC BY 4.0, and carries the minimal rai: fields plus prov:wasDerivedFrom and prov:wasGeneratedBy. It validates with zero errors and zero warnings under mlcroissant 1.1.0:

python -c "from mlcroissant import Dataset; Dataset(jsonld='croissant_metadata.json'); print('ok')"

License

See LICENSE (CC BY 4.0 for this dataset release).