Research map / Multi-agent systems and games / Human-AI and goal-conditioned
EMBGuard: Constructing Hazard-Aware Guardrails for Safe Planning in Embodied Agents
Dongwook Choi, Taeyoon Kwon, Bogyung Jeong, Minju Kim, Yeonjun Hwang, Hyojun Kim, Byungchul Kim, Young Kyun Jang, Jinyoung Yeo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune Qwen-3-VL-2B on EMBHazard (EMBGuard-2B) | Qwen-3-VL-2B · EMBGuardTest | Potential Risk accuracy: 47.2 -> 51.6 | Improved |
| fine-tune Qwen-3-VL-4B on EMBHazard (EMBGuard-4B) | Qwen-3-VL-4B · EMBGuardTest | Potential Risk accuracy: 47.3 -> 54.3 | Improved |
| fine-tune Qwen-3-VL-4B on EMBHazard (EMBGuard-4B) | Qwen-3-VL-4B · EMBGuardTest | Risk Type accuracy: 51.0 -> 50.3 | No clear effect |
| fine-tune Qwen-3-VL-2B on EMBHazard (EMBGuard-2B) | Qwen-3-VL-2B · Held-out Set | Risk Type accuracy: 32.5 -> 59.5 | Improved |
| fine-tune Qwen-3-VL-4B on EMBHazard (EMBGuard-4B) | Qwen-3-VL-4B · Held-out Set | Potential Risk accuracy: 58.3 -> 71.2 | Improved |
| fine-tune InternVL3.5-1B on EMBHazard | InternVL3.5-1B · EMBGuardTest | Potential Risk accuracy gain: +37.6 | Improved |
| fine-tune Gemma-3-4B on EMBHazard | Gemma-3-4B · EMBGuardTest | Potential Risk accuracy gain: +57.3 | Improved |
| scale guardrail from 2B to 4B | EMBGuard · Held-out Set | Hazard accuracy: 36.6 -> 50.1 | Improved |
| scale guardrail from 2B to 4B | EMBGuard · IS-Bench (pre-type process safety) | F1: 30.5 -> 38.3 | Improved |
| provide correct risk type and hazard to policy model | policy models (GPT-4o, GPT-4o-mini, Qwen-3-VL-32B) · IS-Bench mitigation alignment | alignment with ground-truth strategy (%): 28.4 -> 90.4 (both wrong vs both correct); 78.5 wrong risk type; 58.6 wrong hazard | Improved |
| use general MLLM as guardrail instead of specialized EMBGuard (false positive on benign scenarios) | Gemini-2.5-Pro · EMBGuardTest benign scenarios | false positive rate: 83.3% of benign scenarios misclassified as risky | Hurt |
| freeze the vision encoder during fine-tuning (ablation) | EMBGuard · preliminary experiments | hazard explanation quality | Improved |
Closest papers by meaning
- Building a Foundational Guardrail for General Agentic Systems via Synthetic Data (ICLR 2026 · similarity 0.91)
- AGENTSAFE: Benchmarking the Safety of Embodied Agents on Hazardous Instructions (CVPR 2026 · similarity 0.91)
- IS-Bench: Evaluating Interactive Safety of VLM-Driven Embodied Agents in Daily Household Tasks (AAAI 2026 · similarity 0.90)
- HAZARD Challenge: Embodied Decision Making in Dynamically Changing Environments (ICLR 2024 · similarity 0.89)
- SENTINEL: A Multi-Level Formal Framework for Safety Evaluation of Foundation Model-based Embodied Agents (NeurIPS 2026 · similarity 0.88)
- LiSA: Lifelong Safety Adaptation via Conservative Policy Induction (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
