STRIDE: Evaluating Spatiotemporal Reasoning in Driving Edge Cases

Overview of STRIDE: data, tasks, and comparison to prior driving VQA

Figure 1. Overview of STRIDE. Five-frame driving context, six task families, and a sharp drop versus prior driving VQA suites (GPT-6-Astra: 85.2% → 45.5% MCQ on nuScenes).

Introduction

Vision-language models look strong on existing driving VQA benchmarks, but those suites often test static recognition rather than the spatiotemporal reasoning needed in edge cases. STRIDE asks whether a model can track objects over time, judge driving significance, and plan ego motion from multi-frame observations.

The benchmark contains 2,350 questions over nuScenes (1,150) and WOD-E2E (1,200), spanning 39 templates in six families: spatial perception, spatial understanding, temporal memory, temporal extrapolation, trajectory prediction, and scene-context awareness. Formats include MCQ, open-ended questions (OEQ), and geometric waypoint regression. The strongest model we evaluated, GPT-6-Astra, reaches only 45.5% MCQ on nuScenes and 41.2% on Waymo — well above random chance (~20%), but far from prior driving VQA numbers.

2,350questions

39templates · 6 families

5frames per question

45.5%best MCQ (nuScenes)

Leaderboard

Primary ranking key: MCQ accuracy. Open-ended (OEQ) and trajectory (ADE / mini-FDE) metrics are reported separately. Missing entries are shown as —.

#	Model	Type	MCQ ↑	OEQ ↑	ADE ↓	mini-FDE ↓
1 🥇	gpt-6-astra	proprietary	45.5	6.41	0.60	0.98
2 🥈	GPT-5.5	proprietary	39.3	6.68	0.65	0.97
3 🥉	gemini-3.6-flash	proprietary	38.9	6.50	0.69	1.01
4	gpt-5.6-sol	proprietary	38.5	5.65	0.69	0.99
5	gemini-3.1-pro-preview	proprietary	37.8	4.53	1.42	1.84
6	Qwen3.6 35B-A3B	VLM	31.6	5.32	7.42	5.45
7	claude-sonnet-5	proprietary	29.4	6.42	0.87	1.30
8	Qwen2.5-VL 3B	VLM	27.4	3.62	—	—
9	Qwen3-VL 30B-A3B-Thinking	VLM	26.8	5.45	7.12	3.42
10	Qwen3-VL 8B-Thinking	VLM	23.4	5.50	8.00	3.80
11	Dolphins	expert	18.5	3.24	8.16	3.81
12	Qwen3.6 27B	VLM	17.5	5.99	5.81	4.06
—	UniAD 2.0	expert (TRJ-5)	—	—	0.56	0.79
*	Random chance	—	19.8	—	—	—

#	Model	Type	MCQ ↑	OEQ ↑	ADE ↓	mini-FDE ↓
1 🥇	gpt-6-astra	proprietary	41.2	6.77	13.64	12.03
2 🥈	gemini-3.1-pro-preview	proprietary	39.7	4.03	14.22	14.21
3 🥉	claude-sonnet-5	proprietary	39.2	7.01	13.16	10.27
4	gemini-3.6-flash	proprietary	38.8	6.20	13.53	11.37
5	GPT-5.5	proprietary	38.7	7.00	13.84	12.57
6	gpt-5.6-sol	proprietary	37.7	6.48	13.81	12.39
7	Qwen2.5-VL 3B	VLM	37.4	2.57	—	—
8	Qwen3.6 35B-A3B	VLM	34.3	5.83	18.55	8.86
9	Qwen3-VL 8B-Thinking	VLM	32.0	4.72	22.36	19.37
10	Qwen3-VL 30B-A3B-Thinking	VLM	31.8	4.98	11.83	6.18
11	Qwen3-VL 8B-Instruct	VLM	30.1	4.96	19.64	12.65
12	Qwen3.6 27B	VLM	28.4	5.87	20.37	19.25
13	Dolphins	expert	22.2	1.86	18.43	8.60
*	Random chance	—	20.3	—	—	—

OEQ is a 0–10 vision-judge score on open-ended answers. Trajectory errors are in meters (nuScenes vs Waymo scales differ). Machine-readable tables live in the repo. To submit a result, open an issue or PR with your *_score_report.json.

Benchmark

Each question uses a five-frame temporal window. nuScenes inputs are 2×3 surrounding-camera grids; Waymo uses CAM_FRONT. When a target object exists, it is marked with a red box on the query frame. Annotations ship in this repository and on Hugging Face; raw images are downloaded from nuScenes and Waymo Open Dataset.

Space

Spatial perception · SP

Observable object properties: relative position, distance, motion direction, lane occupancy.

Space

Spatial understanding · SU

Driving significance of spatial relations: drivable area, lane change, ego constraint, risk.

Time

Temporal memory · TM

How the current state formed: prior visibility, occlusion, reappearance, recent motion.

Time

Temporal extrapolation · TE

Likely evolution: future motion, occupancy, collision timing, subsequent events.

Context

Trajectory prediction · TRJ

Ego past/future motion via ranking, language justifications, and waypoint regression.

Context

Scene-context awareness · SC

Environment, road structure, and surrounding activity that frame individual objects.

Setup and evaluation instructions: github.com/PlusLabNLP/STRIDE.

Statistics

STRIDE covers challenging driving situations (pedestrians, intersections, construction, debris) across both source datasets, with object-centric questions spanning distance and bearing relative to the ego vehicle.

Scenario, task-family, and ego-behavior coverage of STRIDE

Figure 2. Dataset coverage: (a) scenario tags on nuScenes vs Waymo, (b) task-family mix, (c) future ego speed change by maneuver.

Spatial coverage heatmap of target objects

Figure 3. Spatial coverage of object-centric questions (bearing × distance).

Figure 4. Scenario coverage across the six task families (multi-label clips).

Citation

@misc{liu2026stride,
  title        = {{STRIDE}: Evaluating Spatiotemporal Reasoning in Driving Edge Cases},
  author       = {Liu, Lieqi and Gao, Rui and Gu, Jia-Chen and Hu, Wenbo and Mo, Zhaobin
                  and Moradipari, Ahmadreza and Ammar, Nejib and Wang, Wei and Peng, Nanyun},
  year         = {2026},
  note         = {Preprint},
  howpublished = {\url{https://github.com/PlusLabNLP/STRIDE}}
}