DynamicCatch

Air–ground cooperative search for a moving target. A UAV and a UGV must find one specific vehicle — a car or a two-wheeler (motorcycle / scooter) — that is moving, given only reference views of it, among look-alike moving distractors. The two agents start apart. The UAV searches from 80 m and nominates a candidate; the UGV must meet the vehicle on the road network (≤ 50 m, line of sight) and confirm it with its own cameras. Because the target keeps moving, a location handed from air to ground goes stale — the benchmark is built to measure exactly that.

This release (v2.1) contains the episode specifications — the 10,000-episode dataset on 10 CARLA maps (7,000 car and 3,000 two-wheeler episodes) and a separate 300-episode evaluation set on an 11th map (210 car and 90 two-wheeler episodes) — the map-disjoint splits, the road graphs and visibility maps they refer to, the colour calibrations, and statistics. Closed-loop evaluation runs in CARLA 0.9.15 (+ AdditionalMaps); rendered key frames and reference trajectories will be added in later versions.

Splits

split	maps	episodes (car / two-wheeler)	use
train	Town01, 02, 04, 05, 06, 10HD, 12	7,000 (4,900 / 2,100)	training; splits.json → val (300 ids, 209 / 91) is the only subset used to tune training-free methods
test	Town03, Town07, Town15 (held-out maps)	3,000 (2,100 / 900)	closed-loop evaluation
eval	Town13 (a separate map, not part of the 10k)	300 (210 / 90)	final evaluation of methods
eval_static	static twins of eval	300	same episodes with every vehicle parked (dynamic-gap control)

No map is shared between train, test and eval. Every dataset map holds 700 car episodes (<Town>_0001–_0700) and 300 two-wheeler episodes (<Town>_TW0001–_TW0300). Each of the two id ranges is balanced on its own: car episodes over motion pattern × appearance level × distractor count (5 × 3 × 2, every prefix within ±1), two-wheeler episodes over motion pattern × appearance level (5 × 3; 300 = 20 complete blocks). val takes a prefix of both ranges on every train map, so it is balanced too. The eval set is generated with exactly the same settings as the dataset, only on its own map, and keeps its 7:3 mix: 210 car episodes (Town13_0001–_0210, 7 complete blocks of motion pattern × appearance level × distractor count, so every combination appears exactly 7 times) and 90 two-wheeler episodes (Town13_TW0001–_TW0090, 6 complete blocks of motion pattern × appearance level). Town03 is also the test town of AGOS (static targets).

Episode settings

UAV	nadir camera 1920 × 1440, 53.13° field of view (2 atan 0.5), 80 m above ground (AGL) → 80 × 60 m footprint; 8 m/s; public waypoint grid 20 m
UGV	6 m/s on the road graph, front camera while driving, 6 views when stopped
starts	separated: the UAV launches ≥ min(150 m, half the region side) from the UGV (median 195 m), and no candidate is in its first image
target and distractors	2 m/s on visible roads; 5 motion patterns (constant, turning, stop-go, branch choices, stop-go + branch), stop times jittered per vehicle from the same distribution (identity-neutral)
distractors	cars: 2 or 3 per episode (50 / 50); two-wheelers: 2; similarity level D2–D4
horizon / budget	180 s simulated time, 80 high-level decisions (UAV + UGV), world paused during model inference
success	UGV freshly stopped ≤ 50 m from the target with ≥ 1800 rendered target pixels in a ground view (a two-wheeler's rider counts as part of it); a wrong confirmation ends the episode

Separated starts. The UGV start of every episode is the generator's (≥ 100 m from every candidate, scaled on small maps); the UAV launch point is a 10 m lattice point inside the search region at least min(150 m, region side / 2) from the UGV (150 m on regions of 300 m or more, the Town13 eval included; 131 m on Town07, 94 m on Town10HD, 91 m on Town02), with every candidate outside the initial footprint of a 90° camera — a stricter condition than the 53° camera needs. On 41 small-map episodes no lattice point clears the 90° footprint, and the actual 53° footprint is used (spec.private.sampling.protocol_v2.footprint_rule). The point is drawn deterministically per episode id.

Targets

category	models	classes (share of the category's targets)
car	22	hatch 51%, SUV 23%, coupe 10%, sedan 8%, van 8%
two-wheeler	4	Kawasaki Ninja 29%, Harley-Davidson Low Rider 29%, Vespa ZX125 27%, Yamaha YZF 16% (sport 45%, cruiser 29%, scooter 27%)

Every two-wheeler carries a rider. From above the rider covers most of the vehicle, so all vehicles of a two-wheeler episode share one rider (spec.private.candidates[*].driver_id): the rider never tells them apart, and the level rule below is the only difference. CARLA's three bicycles are not used: their paint covers 4–9 % of their visible body (about 10 px at UGV confirmation range, almost nothing from 80 m), so a colour-defined identity would not be perceivable, and no rider is offered by all three models.

Appearance levels

Levels are defined on rendered colour, calibrated per model and paint in CARLA (assets/render_color_calib_v2.json for cars: median of the central body band; assets/render_color_calib_tw.json for two-wheelers: median of the painted body pixels — the pixels whose colour changes with the paint attribute — with the rider excluded). Each episode's vehicles form an exchangeable set: the level rule holds for every pair of vehicles, and the target is drawn uniformly from the set. Without looking at the target's reference views, no property of the set (colour medoid, odd model out, distance to the UGV or UAV, route) identifies the target better than chance — see the audit below.

level	every pair of vehicles	median rendered ΔE00 to the target, aerial / side: car	two-wheeler
D2	same class, different models, near-identical colour (ΔE00 < 12 in both views; shape only)	5.1 / 7.8	8.6 / 6.9
D3	same model, clearly different shades	23.5 / 25.9	26.1 / 27.2
D4	same model, close shades	14.3 / 15.6	15.7 / 16.1

D2 needs 3–4 different models of one class whose rendered paints match within ΔE00 < 12 in both views. For cars this is possible almost only for hatchbacks and SUVs, so at D2 the car body class is not balanced. For two-wheelers it is possible for the Harley, the Ninja and the Vespa (the YZF's rendered paint never comes within 12 of the Ninja's or the Vespa's), which is also why two-wheeler episodes have 2 distractors: no four motorcycle models match.

Search region: nominally 300 m × 300 m around the start. Small maps shrink it (Town02 181 m, Town10HD 189 m, Town07 261 m); on highway-like maps (Town06, Town15) the box grows to 360–420 m when no closed loop of visible road fits in 300 m.

Per-map statistics (medians unless noted)

map	split	episodes (car / two-wheeler)	region side (m)	UAV–UGV start (m)	UAV–target (m)	UGV–target road (m)	oracle intercept T* (s)	target junction decisions	target stop time (s, mean)
Town01	train	1000 (700 / 300)	299.4	205.9	183.7	385.2	39.0	4.0	4.4
Town02	train	1000 (700 / 300)	181.1	116.2	138.2	271.3	23.0	5.0	4.4
Town03	test	1000 (700 / 300)	300.0	202.5	190.9	350.0	34.0	5.0	4.4
Town04	train	1000 (700 / 300)	300.0	218.4	220.0	239.4	25.0	5.0	4.4
Town05	train	1000 (700 / 300)	300.0	207.5	201.3	487.3	33.0	5.0	4.4
Town06	train	1000 (700 / 300)	360.0	245.0	221.7	246.9	46.0	0.0	4.4
Town07	test	1000 (700 / 300)	261.4	180.7	169.8	280.5	27.0	6.0	4.4
Town10HD	train	1000 (700 / 300)	188.9	121.6	143.7	432.0	31.0	3.0	4.4
Town12	train	1000 (700 / 300)	300.0	219.2	202.0	428.7	45.0	3.0	4.4
Town15	test	1000 (700 / 300)	360.0	216.7	199.8	315.0	38.0	2.0	4.4
Town13	eval	300 (210 / 90)	360.0	240.4	212.0	393.6	52.0	2.0	4.4

Town13 is a large map with long, sparse roads: its regions grow beyond 300 m more often and the targets start farther away, so the eval set is harder than the test maps (median T* 52 s vs 33 s). Car and two-wheeler episodes have the same layout statistics (median T* 33 s vs 34 s, UAV–target 178 m for both).

T* is the earliest time a UGV could intercept the target if its future route were known (a privileged lower bound). Per-episode values are in stats/episodes.jsonl; generation diagnostics in stats/generation_report_*.md.

Shortcut audit

Reference-free heuristics pick one vehicle per episode without seeing the target views (stats/leakage.md, all 10,000 episodes, per category). Chance is the mean of 1 / (number of vehicles): 0.292 for cars, 0.333 for two-wheelers. Every heuristic stays at chance (42 tests; largest excess z = 2.3, i.e. no heuristic is significantly above chance):

heuristic	car D2	car D3	car D4	two-wheeler D2	two-wheeler D3	two-wheeler D4
colour medoid, side view	0.282	0.304	0.282	0.328	0.302	0.329
colour medoid, aerial view	0.276	0.313	0.301	0.323	0.291	0.321
nearest to the UGV start	0.304	0.298	0.299	0.338	0.335	0.321
nearest to the UAV start	0.286	0.286	0.294	0.333	0.336	0.332
longest route / most junctions	0.293 / 0.292	0.291 / 0.286	0.296 / 0.285	0.351 / 0.335	0.330 / 0.340	0.332 / 0.359

"Odd model out" never applies: at D2 all models differ, at D3/D4 all are the same model; "odd rider out" never applies either (one rider per two-wheeler set). The eval set passes the same audit (stats/eval/leakage.md, 300 episodes, 42 tests, every heuristic within |z| ≤ 2.2, largest excess z = 1.7).

Files

data/train.jsonl, data/test.jsonl                                the 10k dataset, one episode per line
data/eval.jsonl, data/eval_static.jsonl                          the separate evaluation map (Town13) and its static twins
splits.json                                                      ids of train / val / test / eval
maps/<Town>/road_graph.json.gz                                   directed road graph (2 m node spacing); node ids used by the specs
maps/<Town>/node_visible.json, node_agl.json, meta.json          aerial visibility per node, ground height, map extent
maps/Town12/*.R{0,1,2}.json, maps/Town13/*.R{0,1,2}.json        per sub-region exports of the large maps
assets/render_color_calib_v2.json                                rendered car colour per (model, paint), 1,104 calibrated pairs
assets/render_color_calib_tw.json                                rendered two-wheeler body colour per (model, paint), 532 pairs, paint-mask sizes, riders
stats/, stats/eval/                                              per-episode statistics, summary, shortcut audit, generation reports

Each row has flat columns for filtering (map, category, motion_group, appearance_level, n_distractors, region_side_m, target_class, target_style, target_model, target_colour, rider (two-wheeler rider id, empty for cars), uav_ugv_m, uav_target_m, ugv_target_road_m, oracle_intercept_s, static_twin) and spec, the full episode as a JSON string:

• spec.public — what an agent may use: map name, search region, UAV / UGV start poses and speeds, the UAV camera (agents.uav.camera) and public waypoint grid (initial_search.uav), horizon.
• spec.private — ground truth for the simulator and the scorer only: which candidate is the target, every vehicle's start, route, stop schedule, model, paint (and rider for two-wheelers), and the privileged intercept witness. Agents must not read private. The target's reference views are rendered from it by the runner.

import json
from datasets import load_dataset

ds = load_dataset("ChipYTY/DynamicCatch")
ep = json.loads(ds["test"][0]["spec"])
print(ep["id"], ep.get("category", "car"), ep["public"]["map_name"], len(ep["private"]["candidates"]) - 1, "distractors")

Simulator and licence

Episodes are specified for CARLA 0.9.15 with the AdditionalMaps package; the road graphs are derived from the maps' OpenDRIVE files. CARLA's assets are licensed CC-BY; please credit CARLA when using this data. This dataset is released under CC BY 4.0.

Changelog

• v2.1 (2026-10-18): the eval set keeps the dataset's 7:3 mix — 210 car episodes (v2.0's Town13_0001–_0210, unchanged) and 90 new two-wheeler episodes (Town13_TW0001–_TW0090, same generator settings, start layouts distinct from the car episodes); Town13_0211–_0300 leave the eval set. eval_static follows. The 10k dataset is unchanged from v2.0.
• v2.0 (2026-10-17): two-wheeler targets and the current evaluation protocol.
  • 3,000 of the 10,000 episodes are two-wheeler episodes (300 per map, ids _TW0001–_TW0300, generated with the same layout / motion code, start layouts distinct from the car episodes of the map); the car episodes are v1.1's _0001–_0700 of every map, unchanged except for the protocol fields. val is re-drawn with the same 7:3 mix.
  • Protocol v2 on every episode (dataset and eval): separated UGV / UAV starts, UAV camera 53.13° (80 × 60 m footprint at 80 m), public waypoint grid 20 m. Only public.agents.uav.{start_pose, grid_spacing_m, camera}, public.initial_search.uav and protocol_version / private.sampling.protocol_v2 change; everything else is byte-identical to the source episode. The eval launch points equal those used in our Town13 evaluation runs.
  • New two-wheeler colour calibration (painted-pixel masks) and per-category shortcut audit.
  • The v1.1 files remain available under the git tag v1.1.
• v1.1 (2026-10-16): the separate evaluation map replaces test_mini: eval (Town13, 300 episodes, same generation settings) and its static twins eval_static. The 10k dataset is unchanged from v1.0.
• v1.0 (2026-10-16): exchangeable appearance sets. v0.1 drew distractors around the target's colour, so the target was the colour medoid of its set (found 53–58% of the time at D2/D4 against 33% chance), and at D2 two distractors could share a model, leaving the target as the odd model out. v0.1 should not be used. Layouts and motion are unchanged.
• v0.1 (2026-10-16): first upload.

Citation

A citation will be added after the review period.