One Sticker Kills Every Robot Policy

🔥 What happened University of Maryland researchers unveiled TAPDreamer, a frozen adversarial patch that drops FastWAM's success rate from 97.7% to 0.0% across 40 LIBERO tasks. The patch covers just 6.5% of the camera input and needs zero queries against the target policy. 💡 Why it matters Prior attacks required constant access to the victim model. TAPDreamer uses only a public encoder and transfers across tasks and architectures — DreamWAM falls to 2.1%, Motus to 10.0%. Random patches of the same size leave 81.5% success intact, proving this is a targeted representation shift, not noise. ⚡ Our take Securing robot policies while leaving the shared vision encoder exposed is a lock without a door. World models need encoder hardening now — before the first physical attack writes the lesson for them.