LG AI Research Unveils Autonomous Vision Inspection Framework Using VLM and Agentic AI

Deep learning vision inspection systems in manufacturing environments suffer from operational bottlenecks because manual labeling and retraining by engineers are essential when Concept Drift occurs. To overcome these limitations, LG AI Research proposed an autonomous vision inspection framework that fundamentally excludes human intervention (Zero-Human Intervention).

This framework is centered on Vision-Language Model (VLM) and Agentic AI, operating through four autonomous mechanisms.

• Uncertainty Estimation and Automatic Sampling: Utilizing Monte Carlo Dropout, it precisely extracts pure uncertainty caused by the model's lack of knowledge (Epistemic), excluding data-specific ambiguity (Aleatoric). This improved accuracy by +4.84% compared to simple random sampling.
• VLM-based Multimodal Automatic Labeling: The VLM infers and labels defect locations and causes in natural language by comparing against normal references. It undergoes training on industrial datasets, domain-specific fine-tuning, and reinforcement learning for inference.
• Noise-Robust Learning: To account for potential errors in pseudo-labels generated by the VLM, a learning policy based on visual feature embedding similarity is applied.
• Agentic AI-based Orchestration: An AI Agent autonomously orchestrates the entire process, from data extraction and retraining to simulation verification and on-site deployment.

As a result, this architecture implements a 'self-evolving data virtuous cycle' that autonomously collects uncertainty data and retrains without manual human intervention.