Researchers demonstrate transferable visual attack on robot control models

Researchers report that a fixed visual patch can defeat multiple robot control models without querying the target policy. Their TAPDreamer preprint describes an attack built with a public visual encoder, which converts camera images into representations used by the control models. Using six frames from one source task, the attack creates patches covering about 6.5% of the input. In closed-loop benchmark tests, it reduces FastWAM success from 97.7% to 0.0% across 40 LIBERO tasks and from 90.8% to 0.0% across 50 RoboTwin tasks. Matched random patches retain 81.5% and 79.2% success. The same attack patches also reduce success on DreamWAM and Motus; the evidence concerns benchmark evaluations.