linyueqian/ME555_llava_v1.5_finetuned

Model Card for Unsolvable Robotic Task Detection

Model Details

• Purpose: Detects when robotic tasks are impossible to complete
• Base Model: LLaVA v1.5 7B
• Developed by: Duke University
• Type: Vision-Language Model

Use Cases

• Identifying unsolvable robotic tasks in real-time
• Explaining why tasks cannot be completed
• Supporting safe human-robot interaction

Training Data

• 4,920 synthetic images with question-answer pairs
• Covers five categories: Status Conflicts, Item Absences, Logical Contradictions, Ambiguous Tasks, and Ethical Constraints

Performance

• Success rate on SDXL synthetic data: 78.05%
• Success rate on simulator synthetic data: 81.00%

Limitations

• Works only with tasks similar to training data
• Requires human oversight
• May not catch novel types of impossible tasks

Getting Started

# Basic configuration
config = {
    "USE_LORA": True,
    "LORA_R": 8,
    "LORA_ALPHA": 8,
    "MODEL_MAX_LEN": 1024
}

Contact

{yixuan.yang,yueqian.lin}@duke.edu