Researchers at MIPT have developed the SG-Safe algorithm, designed to improve navigation for warehouse robots and autonomous vehicles. The system breaks down complex navigation tasks into smaller, manageable stages by generating intermediate sub-goals, which helps the machine accumulate safe movement experience during training.
In computer simulations, including the POLAMP test involving narrow corridors and sharp turns, the algorithm achieved its target in 90% of cases. The system demonstrated an accident rate of 3%, which is 17 times lower than that of existing top-tier approaches.
The algorithm operates significantly faster than traditional methods that calculate routes in real-time. While standard approaches typically require 10 to 30 seconds per step to plan a path, SG-Safe makes decisions more than 10 times faster because it does not need to build routes on the fly.
The research, published in the journal Technologies, highlights that the guidance provided by the system is used only during the training phase. Once training is complete, the robot retains a refined safety program, allowing for near-instantaneous decision-making in real-world scenarios.
While the technology shows high efficiency in simulations, the team noted that it does not yet provide formal safety guarantees for all conditions. Future development plans include simplifying the system architecture and testing the technology on physical robots.
