Motional Open-Sources Dataset to Help Autonomous Vehicles Master Human-Like Reasoning

Motional has released nuReasoning, which it calls the world's first reasoning-centric, long-tail scenario open dataset for autonomous vehicles. The dataset contains 20,000 scenarios drawn from Motional's millions of miles of driving data and is designed to train Vision-Language-Action (VLA) models to understand spatial relationships, anticipate risks, and reason about driving decisions. Each scenario is meant to teach not just the correct action but the logic behind it, helping end-to-end autonomous systems navigate complex edge cases with common-sense intuition. A nuReasoning miniset released earlier this year was downloaded over 50,000 times, signaling strong demand from AI researchers.

Each annotation is human-verified and explains what the vehicle perceives and why it acted as it did. In one scenario from a nighttime construction zone, the AV stops before proceeding onward; the annotation reveals that the stop was correct because a small animal was crossing the road ahead, and alternate routes were rejected because of construction barriers and the animal's unknown speed and direction. This lets researchers see not only what the AV perceived, but why specific actions were taken and why alternative choices were ruled unsafe.

To help researchers navigate the corpus, Motional has integrated its proprietary Omnitag data search engine into nuReasoning. Omnitag allows users to query dataset distributions by scenario type, difficulty level, and location, and it supports natural language semantic search—for example, searching "emergency vehicle approaching behind with a construction site nearby" instantly isolates matching data. The integration is intended to make it easier for researchers to find and examine relevant long-tail scenarios.

Motional and the UCLA Mobility Lab, along with Professor Jiaqi Ma, created the nuReasoning dataset and are hosting the nuReasoning Challenge, launching at the European Conference on Computer Vision (ECCV) in Sweden. The Challenge evaluates planning and reasoning on 1,000 private-test scenarios across two tracks: Explainable Trajectory & Motion Planning, which benchmarks physical motion planning accuracy and safety compliance when guided by explicit counterfactual reasoning, and Long-Tail Visual Question Answering & Scene Reasoning, which evaluates a model's ability to infer spatial relationships, causal decision traces, and risk factors in edge cases. Challenge winners will be announced at the Conference on Neural Information Processing Systems (NeurIPS) in December.

nuReasoning extends Motional's history of open AV datasets, which began with nuScenes in 2019, the industry's first multi-modal public AV dataset, and continued with nuImages, Panoptic nuScenes, and nuPlan. According to Motional, providing open benchmarks lets academic and commercial engineers establish unified safety standards for AV perception, trajectory planning, and explainable AI. Motional's datasets are available for commercial licensing or free academic use under non-commercial terms.