DeliveryGym study shows gains in simulated courier planning

Researchers introduced DeliveryGym, a three-dimensional environment that trains AI agents to plan across a full courier shift, where one delivery can use time, energy or money needed for later orders. In tests across six models and 13 city maps, they found a gap between completing assigned deliveries and choosing the most useful sequence of work. On unseen city maps, reinforcement learning increased Qwen3-VL-4B's net income by 54.3%. Adapting practice shifts to the agent's weaknesses raised evaluation income by 18% over uniform sampling after 100 updates. Both gains were measured in the simulated environment.