Latency-Aware Safety Arbitration for Distributionally Robust Autonomous Driving: A Closed-Loop CARLA Evaluation
Year: 2026
Type: dataset
Abstract: Closed-Loop Benchmark Dataset for Asynchronous Safety Arbitration in Latency-Constrained Distributionally Robust Autonomy (Version 1.0) This repository provides the complete closed-loop CARLA (Town03) simulation benchmark dataset (2,150 simulated episodes and 427,138 total 20 Hz simulation tick records across 8 UTF-8 CSV files) evaluating Dynamic Risk Assessment (FULL_DRA) and its deterministic, latency-aware kinematic Safety Arbiter (DRA_SAFETY_ARBITER) against time-to-collision (BASELINE_TTC) and utilitarian (BASELINE_UTILITARIAN) controllers. EXPERIMENTAL SUITES1. s13_latency_sweep/ (500 episodes; 50 paired cells per latency):Evaluates all four controllers (BASELINE_TTC, BASELINE_UTILITARIAN, FULL_DRA, DRA_SAFETY_ARBITER) across tight occluded-pedestrian corridor scenarios (S13A–S13E, 10 deterministic seeds each) under injected planning latencies L in {0, 75, 150, 300} ms, plus an unguarded FULL_DRA diagnostic sanity check at L = 300 ms. 2. benchmark_suite/ (1,650 episodes; 270 paired cells per latency):Evaluates multi-scenario generalization across 11 diverse urban conflict scenarios (S01–S07, S09–S12; 270 unique scenario–seed cells) at L in {0, 75} ms, plus an additional DRA_SAFETY_ARBITER evaluation at L = 150 ms (n = 270). ARCHIVE STRUCTUREdra_sa_benchmark_v1/ - README.md (Full documentation and reference results) - FallBack_DataSet_Descriptor.pdf (Complete technical descriptor) - s13_latency_sweep/ * s13_fallback_episodes.csv (500 rows, 46 columns — episode-level summary & arbiter telemetry) * s13_episode_results.csv (500 rows, 33 columns — planner-agnostic trajectory metrics) * s13_arbiter_timestep_log.csv (28,000 rows, 21 columns — 20 Hz supervisory ticks) * s13_vehicle_timestep_log.csv (70,000 rows, 24 columns — 20 Hz vehicle kinematics & risk trace) - benchmark_suite/ * benchmark_fallback_episodes.csv (1,650 rows, 46 columns — episode-level summary & arbiter telemetry) * benchmark_episode_results.csv (1,650 rows, 33 columns — planner-agnostic trajectory metrics) * benchmark_arbiter_timestep_log.csv (80,400 rows, 21 columns — 20 Hz supervisory ticks) * benchmark_vehicle_timestep_log.csv (248,738 rows, 24 columns — 20 Hz vehicle kinematics & risk trace) REPRODUCIBILITY & STATISTICAL USAGE GUIDELINES1. Strict Paired-Cell Discipline: All four controllers share identical initial conditions, pedestrian trigger timings, and random seeds within each (scenario, seed, latency_ms) cell. Always perform paired statistical comparisons (e.g., exact McNemar's test for binary outcomes and paired Wilcoxon signed-rank tests for continuous metrics) over matched cells (n = 50 per latency in s13_latency_sweep; n = 270 per latency in benchmark_suite) rather than pooling across heterogeneous latency levels or unpaired runs.2. Treatment of Exploratory Arms: The unguarded FULL_DRA run at L = 300 ms in s13_latency_sweep (n = 50) and the DRA_SAFETY_ARBITER run at L = 150 ms in benchmark_suite (n = 270) are single-arm stress tests rather than complete four-planner paired blocks. Report these runs separately as diagnostic benchmarks rather than merging them into four-way paired hypothesis tests. FUNDING & ACKNOWLEDGMENTSThis work was supported by the African Engineering and Technology Network (Afretec) research grant: "Ethical, Trustworthy, Autonomous: The Vehicles of Tomorrow."
Authors: Zayed, Gamal, El Mougy, Amr
Institution: American University in Cairo
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235213416
