SafeDriveVLA: world model 'dreaming' guards driving VLAs against unsafe navigation commands

UCI6055 · hf · 2026-10-15

• Researchers from UC Irvine, UC San Diego and others propose SafeDriveVLA, the first driving VLA addressing both navigation conflicts and adversarial instructions: free-form natural-language navigation lets distracted users issue unsafe commands — or malicious ones on purpose.
• They also release the first benchmark suite jointly evaluating navigation following, conflict awareness, and adversarial robustness. Existing models fall short: SimLingo completes no language-instructed lane change in closed loop, and all models degrade sharply under conflicting signals.
• Method: SafeDriveVLA explicitly predicts a three-level navigation-compliance mode to decouple conflict reasoning from action generation, and uses navigation-conditioned world model "dreaming" — imagining strictly following the navigation signal — to judge when following is safe.
• Results: 88.16 Driving Score and 68.64% success rate on Bench2Drive; closed-loop language following up 49.3%, and collisions under unsafe instructions reduced from 60 to 38 versus the best SimLingo variant.

Original post →

More from Embodied

• SimDist at RSS2026 lifts failing robot policies to ~90% success with 15-30 minutes of data — KyleMorgenstein · 2026-10-15
• AnyRank: Sequential Testing Ranks Robot Policies Reliably With Up to 50% Fewer Trials — ZeYanjie · 2026-10-15
• Natura-1 Robot Model Demo: Zero-Shot Generalization With Autonomous Retries in One Take — ZeYanjie · 2026-10-15
• Horizon Robotics' VITAL hits 90.8 PDMS with a single denoising step at 3.29x inference speed — HorizonRobotics · 2026-10-15
• Grain-levelling robot keeps workers out of ship holds, cutting engulfment and fall risks — LexiLove · 2026-10-15
• Pi-0.5 can post-train to grab a pen, but legs-balance coordination is the hard part — Darpinian · 2026-10-15

Embodied channel →