Specialists from the Moscow Institute of Physics and Technology (MIPT) have developed the SG-Safe algorithm, which reduces the accident rate of drones and unmanned vehicles by 17 times. This is reported by "Izvestia". The system breaks the route into intermediate sub-goals, so machines do not have to slow down to choose a path.

It is based on reinforcement learning. The algorithm makes the robot "bolder" during training, breaking down complex navigation into short stages. Two strategies: one offers intermediate landmarks, the second is responsible for safe behavior. Hints only work during the training phase; then, a refined program remains in control. Tests showed: the goal is achieved in 90% of cases with an accident rate of only 3%. Decisions are made ten or more times faster – algorithms that build a route on the fly spend 10–30 seconds per step.

First and foremost, the technology is in demand in enclosed areas – warehouses, parking lots, narrow courtyards. After data accumulation, it can be adapted for public roads. Experts note: for warehouses, the indicators are already close to operational, but for automotive transport, the requirements are orders of magnitude stricter, and the transition from simulation to reality almost always leads to a drop in quality.