Hyundai Motor Group described two AI systems in its factories in a company journal article published on 12 August 2026, the day of its AX Achievement Showcase at its Yangjae headquarters in Seoul.
On Hyundai and Kia production lines, each vehicle's specifications and production data are managed by its vehicle identification number (VIN), and every process must confirm that the physical vehicle matches the information registered in the system. Workers used to judge this from the system's specification data or by eye, so on complex lines a mismatch occasionally slipped through and created a risk of fitting the wrong parts. The group first applied automatic VIN recognition on a small scale at some Korean plants, training an AI model on each process's own data and running it on separate servers. As the number of sites grew, managing many models and servers for each plant and line became difficult and maintenance costs rose.
Kwon Oh-hyun of the Vision AI Analytics team restructured the service. The team trained a single model on integrated data from all plants that can run on any infrastructure, moved inference to Amazon Web Services and built a new API that handles large numbers of simultaneous requests without delay. The model recognises each vehicle's VIN and the result is checked against the registered vehicle information in real time, so mismatches are flagged immediately and each process's status can be monitored live. The service now runs at about 70 operating points in all of the group's Korean plants and at plants in the United States, including HMGMA, Europe, India and Asia-Pacific, and the group reports annual savings of about KRW 5.24 billion. It plans to use the same infrastructure to offer image recognition models for other quality inspection tasks.
The second system restores the order of transfer carts that supply parts to the assembly line. Carts move in one direction and can change direction only at fixed points, so queues must be kept in a stable order; when equipment errors or system logic disturb it, the line has to stop and workers had to work out how to move the carts back, which took 10 to 50 minutes depending on their experience and the line. Park Cheol-ung of the Optimization Analysis team treated this as a pathfinding problem and combined the existing search algorithm with reinforcement learning so the system finds the shortest sequence of moves itself. In the side-complete process, which joins the body side panels, it handled 28 carts, more than twice the 12 used during development, where a disordered queue can have millions of possible arrangements, while respecting one-way movement and branching at entry and exit points.
After optimising the program and tuning the search, the system finds the optimal route within the time limit 99% of the time and completes the calculation in 3 seconds on average and 7 seconds at most; it was validated in an environment that replicates the assembly line at Kia's Hwaseong Plant 2. In production, unnecessary downtime fell by about 86%, saving about KRW 40 million a year for a single process, and the team plans to extend it to more plants and processes.