SafeDriveVLA: world-model dreaming catches unsafe navigation commands in driving VLAs
rsasaki0109 · x · 2026-10-13
SafeDriveVLA, a CoRL 2026 paper from UC Irvine, UC San Diego, USC, U. Cincinnati, and NUS, is now open-sourced (Apache-2.0).
The work tackles a new risk: VLA end-to-end driving models accept free-form natural language navigation — but what if the instruction itself is unsafe?
- New benchmarks evaluate instruction following, scene conflicts, and adversarial robustness
- The method combines explicit driving modes with world-model prediction, simulating maneuvers before acting to catch conflicts
- The repo ships benchmark, world model, and training scripts
Related event: SafeDriveVLA Uses World-Model Rehearsal to Block Unsafe Driving Commands(2 posts)→
More from Embodied
- Free-Init: scan-free, motion-free initialization for Doppler LiDAR-inertial systems — rsasaki0109 · 2026-10-13
- On-device robot practices soccer throw-ins with no external tracking — chris_j_paxton · 2026-10-13
- GPT-6 Astra and Claude Opus 5 nearly tied on embodied task: 34-6 over 40 votes, 16-point Elo gap — qinzytech · 2026-10-13
- Custom ComfyUI node repurposes Mixamo's 3D rigs to pose 2D characters — Willonfire8 · 2026-10-13
- First Origami Challenge at IROS 2026: 14 teams push dexterous manipulation to the limit — chris_j_paxton · 2026-10-13
- Robotic hand learns to crawl by treating its fingers as legs — LukeW · 2026-10-13