Ursulalala/HomeGuard-8B
HomeGuard-8B is an 8 billion parameter vision-language safeguard model developed by Ursulalala, built upon Qwen3-VL-8B-Thinking. It is specifically optimized for identifying contextual risks in household tasks for embodied agents. This model excels at grounded multimodal safety reasoning, detecting subtle hazards from environmental context rather than explicit malicious intent. Its primary use case is enhancing safety assessment and planning for embodied agents in domestic environments.
Loading preview...
HomeGuard-8B: Vision-Language Safeguard Model
HomeGuard-8B is an 8 billion parameter vision-language model (VLM) developed by Ursulalala, designed to function as a safeguard for embodied agents. Built on Qwen3-VL-8B-Thinking, this model is uniquely specialized for identifying contextual risks during household tasks, focusing on implicit hazards arising from environmental conditions, object attributes, or spatial relations.
Key Capabilities
- Contextual Risk Identification: Detects subtle, non-explicit safety risks in domestic environments.
- Grounded Multimodal Safety Reasoning: Utilizes visual and linguistic input to reason about potential dangers.
- Safety-Aware Support: Provides crucial safety insights for downstream planning and trajectory generation in robotics.
Training and Specialization
The model was fine-tuned using a step-level Reinforcement Fine-Tuning (RFT) and GRPO-style optimization within the HomeGuard pipeline, leveraging the HomeSafe dataset. This specialized training differentiates it from generic VLMs by focusing on the nuances of household safety.
Intended Use Cases
HomeGuard-8B is ideal for research and development in:
- Safety assessment for embodied agents.
- Contextual risk identification in household tasks.
- Grounded VLM reasoning with visual context.
- Developing safe planning and robotics pipelines.
For more details, refer to the HomeGuard paper and the project repository.