HomeGuard-8B

HomeGuard-8B is an 8B-parameter vision-language safeguard model for identifying contextual risk in household tasks. It is introduced in the paper HomeGuard: VLM-based Embodied Safeguard for Identifying Contextual Risk in Household Task and is designed to help embodied agents detect subtle, implicit hazards that arise from environmental context rather than explicit malicious intent.

This checkpoint corresponds to the 8B step-RFT model used in the HomeGuard framework. It is built on top of Qwen3-VL-8B-Thinking and further optimized for grounded household risk reasoning with reinforcement fine-tuning.

Model Summary

HomeGuard focuses on scenarios where a seemingly benign instruction becomes unsafe because of object attributes, spatial relations, or latent environmental conditions.

Compared with generic VLMs, HomeGuard is specialized for:

• contextual risk identification
• grounded multimodal safety reasoning
• safety-aware support for downstream planning and trajectory generation

Training Recipe

This model is derived from Qwen3-VL-8B-Thinking and trained within the HomeGuard pipeline.

Training setup summarized from the released training configuration:

• Base model: Qwen/Qwen3-VL-8B-Thinking
• Training stage: step-level RFT + GRPO-style optimization in the HomeGuard pipeline
• Training data: HomeSafe

Intended Use

HomeGuard-8B is intended for research and development on:

• safety assessment for embodied agents
• contextual risk identification in household tasks
• grounded VLM reasoning with visual context
• safe planning and downstream robotics pipelines

Usage

This repository contains the inference-ready model weights and tokenizer assets. A typical Transformers loading pattern is:

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model_id = "Ursulalala/HomeGuard-8B"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

For full prompting, evaluation, and application examples, please refer to the HomeGuard project repository.

Resources

• Paper: HomeGuard: VLM-based Embodied Safeguard for Identifying Contextual Risk in Household Task
• Code: AI45Lab/HomeGuard
• Dataset: Ursulalala/HomeSafe
• Base model: Qwen/Qwen3-VL-8B-Thinking

Citation

If you use this model, please cite the HomeGuard paper:

@article{lu2026homeguard,
  title={HomeGuard: VLM-based Embodied Safeguard for Identifying Contextual Risk in Household Task},
  author={Lu, Xiaoya and Zhou, Yijin and Chen, Zeren and Wang, Ruocheng and Sima, Bingrui and Zhou, Enshen and Sheng, Lu and Liu, Dongrui and Shao, Jing},
  journal={arXiv preprint arXiv:2603.14367},
  year={2026}
}