SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving
SafeDriveVLA reduced collisions under unsafe instructions from 60 to 38 compared with the best SimLingo variant in the authors’ tests.
The system imagines what would happen if it strictly followed a navigation instruction, then uses that prediction to assess whether compliance is safe. It separates conflict reasoning from action generation through a three-level navigation-compliance prediction. The authors report a 49.3% improvement in closed-loop language following, alongside an 88.16 Driving Score and 68.64% success rate on Bench2Drive.
HF Daily Papers' note
The system imagines what would happen if it strictly followed a navigation instruction, then uses that prediction to assess whether compliance is safe. It separates conflict reasoning from action generation through a three-level navigation-compliance prediction. The authors report a 49.3% improvement in closed-loop language following, alongside an 88.16 Driving Score and 68.64% success rate on Bench2Drive.
HF Daily Papers' note
score 4