USC paper shows VLAs can keep the same rationale while adversarial perturbations change their actions
USC · hf · 2026-11-05
USC researchers show vision-language-action models can output unchanged reasoning text while visual adversarial perturbations flip the predicted action (e.g. "stop" → "keep going"). The perturbation shifts the shared hidden state in a text-neutral, action-sensitive way, so rationale-only monitors miss it — a key safety warning for robot and driving VLA deployment, with diagnostics and steering methods proposed.
More from Embodied
- "Robotics simulation platform" startups are a dead end — agents already removed the need — carlosdponx · 2026-11-05
- Runtime (YC F26) debuts simulation-to-real robot policies, threads a rope with no fine-tuning — ycombinator · 2026-11-05
- Rooms3d turns Vision Pro room scans into permanent 3D dioramas for $9.99, one-time purchase — Scobleizer · 2026-11-05
- Robot world model demo shows a grasp that never happened in reality — Negative-Whereas3307 · 2026-11-05
- China demos AI blackboard turning sketches into live 3D models, claimed in 60,000 schools — CurieuxExplorer · 2026-11-05
- Gym-Anything for Robotics uses coding agents to mass-generate realistic robot simulations — wellecks · 2026-11-05