ylee-sling/aero-vpa-benchmark

Aero-VPA closed-loop benchmark (PX4 SITL, procedurally generated forests)

Every closed-loop mission flown to evaluate the Aero-VPA policies (weights): one record per mission, decision-by-decision traces of the traced runs, the worlds with their privileged obstacle maps, and the reports. Generated by scripts/release/package_release.py (2026-11-03, code c420bf6).

Contents

path	contents
missions/missions.parquet	one row per mission of every run (2448 rows): run, tag, world, density, episode, ending, success, duration, path, speeds, clearance, waypoints, fallbacks, blind decisions, shield interventions
records/	the raw records written by sim/fly_policy.py (JSON lines), one file per run and simulator
records/dagger/	the DAgger rounds' supervised flights (expert-gated, not autonomous performance)
traces/<set>/	decision-by-decision traces (fly_policy.py --trace), gzipped JSON lines; the first line is the mission record
worlds/	the worlds flown (Gazebo SDF) and their privileged obstacle maps (JSON)
reports/	text reports of the analysis scripts

Protocol

36 test worlds (seeds 2000-2011 × sparse, medium, dense; Poisson-disc trunk spacing 7, 5 and 3.5 m), 5 missions each: the benchmark runs fly 180 missions; pilot and traced runs fly mission 0 of every world, and earlier baselines subsets of the worlds (run and tag identify each). Mission k of a world is seeded by (world seed, k), so every run flies the same missions. A mission takes off to 2.5 m and visits 3-6 waypoints 10-40 m apart; it times out at 1.5 × the expert reference (3 m/s) + 15 s. Endings: success, collision (privileged distance below 0.35 m), ground (below 0.3 m), timeout, camera_timeout (no decision for 0.5 s), out_of_bounds. Shielded runs fly with fly_policy.py --shield (onboard depth memory, barrier, blind-motion cap, altitude floor, breadcrumb backoff; no map).

Results

checkpoint	success, no shield	success, shield [95% CI]	crashes, no shield → shield	timeouts, shield	paired McNemar p (success)
r1	38/180 (21.1%)	94/180 (52.2%) [45.0, 59.4]	72.2% → 22.2%	23.3%	1.20 × 10-10
kw	28/180 (15.6%)	90/180 (50.0%) [42.8, 57.2]	80.6% → 26.1%	21.7%	4.69 × 10-12
sd	28/180 (15.6%)	82/180 (45.6%) [38.4, 52.8]	76.1% → 35.0%	16.1%	4.18 × 10-11
nomotion	24/180 (13.3%)	81/180 (45.0%) [37.9, 52.3]	77.2% → 42.8%	10.0%	1.33 × 10-10
r2	23/180 (12.8%)	77/180 (42.8%) [35.8, 50.1]	80.0% → 38.3%	16.7%	4.18 × 10-11
noact	20/180 (11.1%)	79/180 (43.9%) [36.8, 51.2]	82.2% → 36.1%	14.4%	2.45 × 10-12

Success by world density (shield / no shield, missions per density in brackets):

checkpoint	dense	medium	sparse
r1	28 / 4 (60)	29 / 17 (60)	37 / 17 (60)
kw	22 / 1 (60)	33 / 10 (60)	35 / 17 (60)
sd	17 / 2 (60)	26 / 9 (60)	39 / 17 (60)
nomotion	17 / 1 (60)	28 / 6 (60)	36 / 17 (60)
r2	21 / 0 (60)	32 / 9 (60)	24 / 14 (60)
noact	23 / 2 (60)	25 / 6 (60)	31 / 12 (60)

Missions the shield was not revised on (mission 1 and later of every world; the shield pilots flew mission 0):

checkpoint	missions	success, no shield	success, shield [95% CI]	paired p
r1	144	21.5%	52.1% [44.0, 60.1]	2.00 × 10-8
kw	144	15.3%	54.2% [46.0, 62.1]	2.28 × 10-12
sd	144	16.7%	45.1% [37.2, 53.3]	5.26 × 10-8
nomotion	144	15.3%	46.5% [38.6, 54.7]	3.74 × 10-8
r2	144	13.9%	44.4% [36.6, 52.6]	2.40 × 10-9
noact	144	10.4%	43.1% [35.3, 51.2]	4.32 × 10-10

Reproduce the numbers

git clone https://github.com/ylee-sling/Aero-VPA && cd Aero-VPA
hf download ylee-sling/aero-vpa-benchmark --repo-type dataset --local-dir bench
python3 sim/paired_stats.py --root bench/records
python3 sim/summarize_flights.py bench/records/flights_shield_r1_i*.jsonl --by density
python3 sim/crash_study.py 'bench/traces/trace_r1/*.jsonl.gz' --world_root bench/worlds/test

import pandas as pd
df = pd.read_parquet("hf://datasets/ylee-sling/aero-vpa-benchmark/missions/missions.parquet")
print(df.groupby("run")["success"].mean())

Limitations

Simulation only; static forests; no wind; one training seed per policy; mission outcomes have a large chance component. Collisions are detected with the privileged map. The shield was revised after its first pilot, which flew mission 0 of every test world.

License

CC BY 4.0. The code that produced and reads these files is GPL-3.0-only: https://github.com/ylee-sling/Aero-VPA. For commercial integration, proprietary licensing, or defense adaptation, contact ylee@sling.ai.kr.