Yandex delivery robots have received a major update: now their movement trajectory is helped by AI based on a neural network with a transformer architecture. Thanks to the new planner, robots better assess the surrounding environment, more accurately predict the behavior of pedestrians, cyclists, scooter riders, and cars, and more confidently handle unusual situations. According to test results, delivery has become 10% faster on average.
Previously, rovers primarily relied on an algorithmic planner – a set of predefined rules. In a complex urban environment, this approach often forced the device to act cautiously: slow down, stop, and wait until the path was completely clear. The new neural network planner chooses a more flexible trajectory: the robot tries to continue moving even in busy places, but without interfering with people and other road users.
For safety, the AI does not work alone. The neural network and algorithmic planners offer their route options, after which the system selects the optimal one. This dual loop helps robots move faster while maintaining caution. Currently, under the control of the neural network, delivery robots spend an average of about 80% of their trip.
Yandex has also updated the surrounding notification system. The robots still have a glowing flag at an adult's eye level, and now they can additionally emit a soft sound signal. It does not require giving way, but simply warns that a robot is moving nearby. The signal is not constantly on, but only in situations where it is important to indicate its presence.
The neural network planner was trained in a simulator where thousands of road scenarios can be quickly played out – from ordinary walks on the sidewalk to unexpected obstacles. The total duration of virtual trips has already exceeded 4000 years: Yandex compares this to training that began in ancient Mesopotamia.
To build the route, robots use data from cameras and lidar, and all calculations are performed directly on board. Therefore, the delivery robot will not go astray even if the internet connection is lost. Additionally, the devices are equipped with ultrasonic sensors that help detect obstacles in time and avoid collisions.