Event

SAGE: Vision-Language and Geometry Fusion for Work Zone Path Prediction

1 reports · 1 sources

Jiaxi Liu and other authors propose the SAGE (Semantic-to-Action Grounded Execution) framework for path prediction in temporary traffic control work zones: it treats trajectories output by a vision-language model (VLM) as preferences over complete geometric candidate paths, then uses ranking fusion to combine the candidates' geometric scores with VLM trajectory consistency. The framework separates path construction, path preference, and control compliance into distinct stages within a shared candidate space. The authors report that among 2,306 real ROADWork validation clips, 87% of clips had a candidate set containing a path within 10 pixels of the human demonstration, but errors in the selection stage accounted for 53% of the geometric ranker's total error. SAGE-VLM is trained with supervision grounded in the Manual on Uniform Traffic Control Devices (MUTCD), requiring it to state lane status and hazard side before giving a passing decision and trajectory.