ADAS programs struggle with consistency: per sensor annotations, calibration drift, and shifting guidelines create fragmented ground truth that slows training and erodes trust. This talk presents a 4D vector space labeling approach that anchors annotations in world space over time, not in individual sensors. The result: one global, spatiotemporally consistent source of truth spanning cameras, radar, ultrasonics, and lidar—while charting a pragmatic path to vision only as sensors are reduced. We’ll connect lessons from large scale deployments to the data layer: how to represent world models, enforce cross modal consistency, and keep labels recomputable as calibrations and models evolve. We'll discuss tooling patterns that make this work: multimodal visualization, search and relabel loops, and observability that surfaces inconsistency. The payoff: simpler operations, lower CapEx over time, and models that learn from one coherent reality instead of a stack of conflicting sensor views.

• The current development status
• Where research and development need to work together more closely
• What the future of development and research could look like