AI TrackerFrontier desk · AEST
Back to news
normalaitech

TAPDreamer: Transferable Adversarial Patches for World Action Models

arXiv cs.AI · 6 Oct 2026, 03:55 AEST

TAPDreamer introduces a novel attack on world action models that utilizes a public encoder to create transferable adversarial patches. This method significantly reduces the success rates of various robotic tasks without needing access to target-policy queries.

Key facts

Read aloud

TAPDreamer introduces a novel attack on world action models that utilizes a public encoder to create transferable adversarial patches. This method significantly reduces the success rates of various robotic tasks without needing access to target-policy queries. World models learn to predict how their environment will evolve, making them an important foundation for general-purpose robotic control. In this paper, we propose an attack, TAPDreamer, against world action models that instead uses a public encoder alone to construct a fixed local perturbation that transfers across tasks and action architectures. In closed-loop evaluation, one frozen patch per benchmark, covering about six point five percent of the input, reduces FastWAM's success rate from ninety seven point seven percent to zero point zero percent across forty LIBERO tasks and from ninety point eight percent to zero point zero percent across fifty RoboTwin tasks. These results show that protecting downstream action generation alone is insufficient: defenses for world action models must also secure shared visual encoders against persistent local perturbations.

Sources in this cluster