Motional Releases Open-Source Dataset to Teach AVs Human-Like Reasoning

Modern autonomous systems are shifting from simple perception to high-level logic, requiring Vision-Language-Action models that understand the "why" behind every maneuver. While current AI can identify objects, it often struggles with the common-sense intuition needed for edge cases—like distinguishing between a construction barrier and a crossing animal. Motional’s new dataset addresses this by offering 105 hours of high-quality, human-verified reasoning annotations, allowing researchers to train models on spatial, decision-based, and counterfactual logic.

The dataset, developed alongside the UCLA Mobility Lab, incorporates data from Motional’s global fleet across cities including Las Vegas, Singapore, and Boston. Researchers can utilize the integrated Omnitag search engine to query scenarios using natural language, such as searching for emergency vehicles near work zones. To accelerate adoption, Motional is hosting the nuReasoning Challenge at the European Conference on Computer Vision in Sweden, with winners to be announced at NeurIPS in December. This release follows the company’s history of open-source contributions, including the 2019 nuScenes project, as it seeks to establish industry-wide benchmarks for safety and explainable AI.