Yandex Delivery Rovers Get 40% Faster With AI Motion Planner

The company relies on reinforcement learning. After each virtual route, the system scores how correctly the rover completed it, and that feedback helps the model pick the safest and most efficient trajectories. Yandex says its models have spent more than 20,000 years in simulation in total. One hour of simulation gives the virtual rovers about 140 days of experience, and 40 hours equals roughly 15 years.

From fixed rules to an AI planner

The whole fleet switched to the new motion control system in May 2026. Before that, the robots followed algorithms with behavior rules written in advance by engineers. In complex or unfamiliar situations they tended to play it safe: slow down and wait for the path to clear.

The AI planner lets rovers keep moving in tougher conditions without getting in the way of others. A rover can, for example, slow down ahead of an obstacle in good time or speed up on a clear stretch.

Safety layers remain

Despite the AI training, several independent safety layers still operate on real routes. One is a lidar loop that reacts to obstacles within a tenth of a second and automatically limits the robot's speed. When an obstacle appears at close range, ultrasonic sensors, built on the same principle as a car's parking sensors, kick in and can bring the rover to a full stop if needed.

For customers, a faster and less hesitant rover should mean smoother deliveries in crowded areas, while the layered safety setup is meant to keep the more assertive behavior in check.