Managing an AGV Fleet with Reinforcement Learning: The PPO Case

The story of moving from rule-based dispatching to reinforcement learning, told through a real case: how line stoppages dropped by 98.6% and fleet utilization rose.

AGV fleets in most facilities are still managed with fixed rules: nearest vehicle, oldest task, or simple priority order. This approach works on a calm floor, but it quickly clogs up under peak load and variable demand.

The Real Shape of the Problem

The floor is dynamic: line demand, AGV position, battery level, traffic, and charger availability change constantly. A rule-based system tries to handle this variability with pre-coded scenarios — but real life has more scenarios than you can code for.

A concrete picture from one production facility: an average of 21 line stoppages per shift, 72% fleet utilization, and an average charging time of 70 minutes. The rule-based system had hit its limits.

Why PPO Fits This Problem

Proximal Policy Optimization (PPO) is a balanced variant of reinforcement learning: it strikes a good balance between exploration and stability, which matters in low-error-tolerance settings like a production environment. The model makes every dispatch decision by weighing AGV positions, battery levels, task priorities, line demand, and traffic data together.

Testing Without Touching the Floor

The decision engine was developed not directly on the floor but in a physics-based simulation and a Unity digital twin. The AGV fleet, lines, task points, and charging stations were modeled virtually; different scenarios were tested without disrupting operations. This mattered both for safety and for development speed.

The Results

After a four-month development and rollout process — see the full case here — line stoppages dropped from ~21 to ~0.3 per shift (a 98.6% reduction), active AGV ratio rose from 72% to 91%, average charging time fell from 70 to 50 minutes, and completed tasks increased by 39% with the same fleet capacity.

The Safety Boundary

This isn't a black box making decisions freely. Whenever a situation falls outside safety and operating rules, the system reverts to the classic, predefined safe decision. AI delivers the optimization while floor safety and operational continuity are preserved; every decision is logged and remains comparable against a baseline.