HomeSafe-Bench: Evaluating Vision-Language Models on Unsafe Action Detection for Embodied Agents in Household Scenarios

Conference: CVPR 2025
arXiv: 2603.11975
Code: Yes
Area: Multimodal VLM / AI Safety
Keywords: Household safety, unsafe action detection, VLM evaluation, embodied AI, dual-brain architecture

TL;DR

HomeSafe-Bench is the first benchmark to evaluate VLMs on unsafe action detection in household scenarios (438 cases covering 6 functional areas), and proposes HD-Guard, a hierarchical streaming architecture that coordinates a lightweight FastBrain and a large-scale SlowBrain to achieve real-time safety monitoring.

Background & Motivation

Background: Household robots are developing rapidly, but household environments introduce unpredictable safety risks (e.g., perception delays, lack of common-sense leading to dangerous operations). Existing safety evaluations are mostly limited to static images, text, or general harms.

Limitations of Prior Work: (1) Lack of a standardized benchmark for dynamic unsafe action detection; (2) household scenarios are more complex and variable than industrial environments, requiring contextual understanding to determine if an action is safe; (3) the capabilities and bottlenecks of VLMs in safety detection remain unclear.

Key Challenge: Real-time safety monitoring demands low latency, whereas accurate unsafe action detection requires deep multimodal reasoning—making it difficult to achieve both simultaneously.

Goal: To build an evaluation benchmark and design a real-time safety monitoring architecture.

Key Insight: (1) Construct a diverse unsafe action dataset through a hybrid pipeline of physical simulation and video generation; (2) utilize a dual-brain architecture to balance inference efficiency and detection accuracy.

Core Idea: FastBrain performs high-frequency lightweight screening, while SlowBrain conducts asynchronous deep reasoning, coordinating both to achieve real-time safety.

Method

Overall Architecture

HomeSafe-Bench contains 438 unsafe cases covering 6 functional areas such as kitchens and living rooms, with multi-dimensional fine-grained annotations. During inference, HD-Guard coordinates the fast and slow dual brains: FastBrain continuously screens video frames at a high frequency, triggering SlowBrain for deep multimodal analysis when suspicious actions are detected.

Key Designs

1. Hybrid Data Construction Pipeline:

   • Function: Generate diverse and realistic videos of unsafe actions
   • Mechanism: A physical simulator generates basic scenes and actions, combined with advanced video generation models to enhance visual realism, followed by human annotation of unsafe categories, severity, and context
   • Design Motivation: Pure simulation lacks realism, while pure real-world data is difficult to scale to cover sufficient unsafe scenarios

2. Hierarchical Dual-Brain Guard (HD-Guard):

   • Function: Real-time safety monitoring architecture
   • Mechanism: FastBrain is a lightweight model (e.g., a small ViT) that scans video frames at high frequency to output quick safety scores for each frame. When a score exceeds a threshold, it asynchronously triggers SlowBrain (a large VLM like GPT-4V) to conduct deep multimodal reasoning, combining vision, language, and common-sense knowledge to make the final decision
   • Design Motivation: Analogous to the human fast-and-slow systems (System 1/2)—fast intuitive judgment suffices most of the time, and deep reasoning is initiated only when necessary

3. Multi-dimensional Fine-grained Annotation:

   • Function: Support systematic evaluation
   • Mechanism: Each case is annotated with the type of unsafety (e.g., collision, fall, fire), severity, involved objects, and contextual dependency. The division into 6 functional areas ensures the evaluation covers various typical spaces in a home
   • Design Motivation: Coarse-grained "safe/unsafe" binary classification is insufficient to diagnose specific weaknesses of the models

Loss & Training

FastBrain can be fine-tuned with a small amount of annotated data, while SlowBrain uses pretrained VLMs for zero/few-shot inference.

Key Experimental Results

Main Results

Method	Detection Accuracy	Latency	Explanation
HD-Guard	Best trade-off	Low	Dual-brain coordination
Large VLM only	Highest accuracy	Very High	Unsuitable for real-time
Lightweight model only	Lower	Lowest	Severe missed detections

Ablation Study

Configuration	Performance	Explanation
FastBrain + SlowBrain	Best	Complementary coordination
FastBrain only	High missed detections	Lacks deep reasoning
SlowBrain only	High latency	Unable to run in real-time
Different trigger thresholds	Trade-off exists	Lower threshold -> More SlowBrain calls

Key Findings

• Existing VLMs perform far from perfectly in unsafe action detection, especially in scenarios requiring common-sense reasoning (e.g., determining whether "a knife pointing towards a toddler" is hazardous).
• The fast-slow dual-brain coordination of HD-Guard significantly outperforms single-model strategies.
• Contextual dependency is a key bottleneck—the exact same action may be safe or unsafe depending on the context.

Highlights & Insights

• Critical Safety Evaluation Gap: This is the first work to systematically evaluate the capability of VLMs in household unsafe action detection, which holds direct significance for embodied AI safety.
• Practicality of the Dual-Brain Architecture: The analogy to System 1/2 is intuitive and effective; this architecture can be transferred to other scenarios requiring real-time monitoring and deep analysis.
• Hybrid Data Construction: The simulation + generation pipeline presents a practical solution to address the scarcity of safety-related data.

Limitations & Future Work

• The scale of 438 cases is relatively small and may not cover all unsafe scenarios.
• Unsafe actions generated via video models may possess a distribution shift compared to real-world actions.
• The optimal balance between the false positive rate of FastBrain and the invocation frequency of SlowBrain requires scenario-specific tuning.
• Safety concerns in multi-person interaction scenarios are not considered.

Related Work & Insights

• vs SafetyBench (Text Safety): Text-based safety evaluations do not involve visual and physical interactions, whereas HomeSafe-Bench is closer to embodied scenarios.
• vs RoboCasa/Habitat: Simulation platforms provide environments but do not focus on safety evaluations; HomeSafe-Bench fills this gap.
• vs System 1/2 Architectures: HD-Guard serves as a concrete implementation of dual-system theory in cognitive science for AI safety.

Rating

• Novelty: ⭐⭐⭐⭐ The combination of a new benchmark and a dual-brain architecture is novel.
• Experimental Thoroughness: ⭐⭐⭐⭐ Includes multi-VLM comparisons and comprehensive ablations.
• Writing Quality: ⭐⭐⭐⭐ The problem motivation and methodology are clearly described.
• Value: ⭐⭐⭐⭐⭐ Holds significant practical importance for embodied AI safety.

Related Papers

• [CVPR 2025] From Multimodal LLMs to Generalist Embodied Agents: Methods and Lessons
• [CVPR 2025] Embodied Scene Understanding for Vision Language Models via MetaVQA
• [CVPR 2025] Evaluating Vision-Language Models as Evaluators in Path Planning
• [ICML 2026] VLA-Arena: An Open-Source Framework for Evaluating Vision-Language-Action Models
• [NeurIPS 2025] MME-VideoOCR: Evaluating OCR-Based Capabilities of Multimodal LLMs in Video Scenarios