Datasets for "Reversal and Divergence in Compressed Airborne Collision Avoidance Policies" (ACAS Xu, HorizontalCAS, VerticalCAS)

This repository contains the datasets behind every result in the paper. It covers three studies: (1) the ACAS Xu study (Sections 5–6), with 30,000 sampled encounters, 480,000 rollout rows across policy arms and intruder conditions, a 73,800-row DAIDALUS cross-check, trained S7 and S7+ surrogate network weights, and simulation configuration with random seeds; (2) the reference-systems study (Section 6) for HorizontalCAS and VerticalCAS, with 720,000 rollout rows, 30,000 VerticalCAS encounter parameters, regenerated-table validation against the released networks, a reversal-penalty sweep (0.1×, 1×, 10×), and Julia cross-checks; and (3) the mechanisms study (Section 7), with 660,000 rollout rows across systems, policy pairs, targeting schemes and conditions, plus recoverability searches and the tables behind Figures 7–13. Aggregated summary tables for the ACAS Xu study are included. Common conventions: distances in feet, times in seconds, angles in radians; eid links rows across files; condition codes are nominal, turn_only, vertical_only, full; advisory codes follow ACAS Xu/HorizontalCAS (COC, WL, WR, SL, SR) and VerticalCAS (COC, DNC, DND, DES1500, CL1500, SDES1500, SCL1500, SDES2500, SCL2500). Large per-encounter files round distances to whole feet; outcome flags (reversal, nmac) are exact. Load with pandas, e.g. pd.read_csv('data/acasxu/rollouts.csv.gz'), or use the helpers in src/acasxu_reversal/data.py. Files regenerable deterministically from the same seeds (per-encounter reversal-penalty sweep rows, mechanism runs targeting the other policy's escape path, solved score tables, full 1 Hz trajectories) are omitted to keep the repository small; the main README scripts regenerate them.