Reproduce “Manipulation Facing Threats: Evaluating Physical Vulnerabilities in End-to-End Vision Language Action Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hao Cheng, Erjia Xiao, Yichi Wang, Chengyuan Yu, Mengshu Sun, Qiang Zhang, Jiahang Cao, Yijie Guo and 7 more

Researchers evaluated the physical vulnerabilities of open-source end-to-end Vision Language Action Models (VLAMs) against visual threats using a new Physical Vulnerability Evaluating Pipeline (PVEP). The study found that models like LLaRA and OpenVLA are highly susceptible to environmental degradations and adversarial patches, with some attacks leading to over 90% task failure rates.

Key insights

  • All categories of physical visual attacks significantly degrade VLAM performance, leading to increased task failure rates and extended task completion times.
  • Adversarial patches generated from base Multimodal Large Language Models (MLLMs) can exhibit adversarial transferability to fine-tuned VLAMs, even with limited attacker knowledge.
  • The vulnerability of VLAMs to physical threats is highly dependent on the type and intensity of the attack, as well as the attacker's knowledge level.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.