Why Strong AI Vision Models Often Fail When Deployed on Factory Floors

Lê Thái Sơn, a project manager at Solomon Technology Corp., argues that a vision model performing well in a demo addresses only part of an industrial problem. He calls the broader challenge industrial AI engineering, which he says differs fundamentally from AI research. His framework treats a robot's perception, cognition, and action as one chain to be evaluated as a whole rather than stage by stage. He notes that real factories bring changing lighting, vibration, dust, product variation, and fluctuating conveyor speeds that lab datasets do not capture. He also stresses system-level questions about camera stability, processing speed, failure handling, and what the system should do when the AI is uncertain.