From 56c0847c02c19a804a690f7ededa2b421c60ed8a Mon Sep 17 00:00:00 2001 From: Xiang Zheng Date: Sun, 24 May 2026 05:16:35 +0800 Subject: [PATCH] add Safety section + Embodied AI Safety Survey MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit WAMs as joint video-action generators / autoregressive predictors / diffusion planners inherit safety vulnerabilities at every embodied layer they integrate (perception fusion, cognitive planning, action control, agentic cascading). The Awesome-WAM list lacks a Safety section to surface this threat surface. Adds: - New top-level `## Safety` section between Evaluation and Citation - TOC entry `- [Safety](#safety)` - First entry: "Safety in Embodied AI: A Survey of Risks, Attacks, and Defenses" (arXiv:2605.02900) — 500+ paper survey across 5 layers / 18 subcategories, structured around Capability-Risk Duality (capability expansion = new vulnerabilities; inner-layer vulnerabilities cascade outward). Single-paragraph blurb explains relevance to WAM design (threat surface inherited by joint/cascaded WAMs at every layer they integrate). Open to PRs from the community to expand this section with additional safety-focused references. --- README.md | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/README.md b/README.md index 7ad954d..2425a7d 100644 --- a/README.md +++ b/README.md @@ -67,6 +67,7 @@ This repository accompanies our survey on **World Action Models (WAMs)** — the - [Evaluation](#evaluation) - [Training Data](#training-data) - [Evaluation](#evaluation) +- [Safety](#safety) ## Tag Legend ### World Action Model tags @@ -630,6 +631,17 @@ This repository accompanies our survey on **World Action Models (WAMs)** — the +## Safety + +Safety considerations across the embodied AI pipeline that World Action Models inherit: perception robustness, cognitive jailbreak, planning manipulation, action-level adversarial control, and agentic cascading risks. + +- **Embodied AI Safety Survey**: "Safety in Embodied AI: A Survey of Risks, Attacks, and Defenses", arXiv 2026. + [[📄 Paper](https://arxiv.org/pdf/2605.02900)] [[🌍 Webpage](https://x-zheng16.github.io/Awesome-Embodied-AI-Safety/)] [[💻 Code](https://github.com/x-zheng16/Awesome-Embodied-AI-Safety)] [[🌟 Blog](https://github.com/x-zheng16/Awesome-Embodied-AI-Safety#-news)] + + First unified safety framework for embodied AI, covering 500+ papers across 5 layers (Perception, Cognition, Planning, Action and Interaction, Agentic) and 18 subcategories. Capability-Risk Duality: each layer of the embodied pipeline represents a capability expansion that introduces corresponding new vulnerabilities; inner-layer vulnerabilities cascade to outer layers. Relevant to WAM design as it characterizes the threat surface that joint video-action generators, autoregressive predictors, and diffusion planners inherit at every layer they integrate. + + + ## 👋 Citation If you find this survey or repository helpful for your research, please consider citing our paper: