MIT researchers have developed a novel AI technique to enhance safety and stability in autonomous robots, addressing the challenging ‘stabilize-avoid’ problem. This method employs a two-step approach combining deep reinforcement learning with mathematical optimization and was successfully tested using a simulated jet aircraft. Unlike existing AI methods struggling with conflicting objectives of stability and obstacle avoidance, this new technique achieves a tenfold increase in stability while maintaining safety, enabling machines to effectively navigate complex scenarios.
The team’s paper, led by graduate student Oswin So and assistant professor Chuchu Fan, explains how the problem was reframed as a constrained optimization issue. The researchers utilized a mathematical representation called the epigraph form, allowing them to leverage reinforcement learning effectively. Their tests demonstrated that this approach could control a simulated jet, ensuring it stayed on course while avoiding collisions, outperforming existing methods.
The implications of this research are significant, particularly for dynamic robots like autonomous delivery drones, which require both safety and stability. Future developments aim to incorporate uncertainty into the optimization process and explore real-world applications of the algorithm in hardware, enhancing its practical utility in mission-critical systems.
