Event
TAPDreamer Uses Fixed Adversarial Patch to Drive World Action Model Success Rate to Zero
1 reports · 1 sources
Xuanyu Lu et al. propose TAPDreamer, a black-box transferable adversarial patch attack method against world action models that uses only public encoders to construct a fixed local patch, requiring no queries to the target policy and transferring across tasks and action architectures. In closed-loop evaluation, with one frozen patch per benchmark covering about 6.5% of the input, FastWAM's success rate dropped from 97.7% to 0.0% on 40 LIBERO tasks and from 90.86% to 0.0% on 50 RoboTwin tasks, while matched random patches still retained 81.5% and 79.2% success rates. The method uses six frames of images from a single source task and maximizes the global L1 distance between encoder representations before and after adding the patch to construct the patch.
Report timeline
- TAPDreamer: Transferable Adversarial Patches for World Action Models
TAPDreamer proposes an attack method that constructs a fixed local adversarial patch using only a public encoder, requiring no queries to the target policy and transferring across tasks and action architectures. In closed-loop evaluation, a frozen patch covering about 6.5% of the input reduced FastWAM's success rate from 97.7% to 0.0% on 40 LIBERO tasks and from 90.86% to 0.0% on 50 RoboTwin tasks, while a matched random patch still retained 81.5% and 79.2%.
arXiv cs.RO ·