Scientists from the Massachusetts Institute of Technology (MIT) have developed a new method that uses machine learning with reinforcement and mathematical optimization to ensure the safety and stability of autonomous robots. The approach was successfully tested on a simulated jet aircraft that could fly through a narrow corridor without encountering the ground.
The task of achieving a goal while avoiding obstacles and preserving trajectory is known as the problem of “stabilize-raising”. Many existing methods of artificial intelligence cannot safely achieve this goal.
The method developed by the MIT team is based on a two-stage approach. First, a robot is trained using a neural network that receives a reward for achieving the goal and a penalty for colliding with obstacles. Then, mathematical optimization is used to clarify the robot’s behavior and guarantee its safety and stability.
The study showed that this approach is more effective than other methods, ensuring the same or better safety with a tenfold increase in stability. The robot reached and remained stable in its target area, even in a simulated scenario that required it to navigate a narrow corridor without colliding with the ground.
Chuchu Fan, Associate Professor of Aeronautics and Astronautics and member of the Information Laboratory and Decision (Lids), explained that the problem of “stabilize-raising” is a long and difficult one as many people studied it, but did not know how to cope with the high complexity of dynamics.
The algorithm developed by the MIT team could be used in dynamic robots that require security and stability, such as autonomous drones for delivering goods. The findings will be presented at the Robotics conference: Science and Systems.
Source: arXiv