MIT researchers have developed a machine learning approach that can address the challenge of stabilizing autonomous vehicles while avoiding obstacles more effectively than previous methods. The new technique, presented in a paper by lead author Oswin So and senior author Chuchu Fan, enables autonomous aircraft to navigate challenging terrain with a tenfold increase in stability and ensure safe operations while achieving their goals.
The researchers reframed the stabilize-avoid problem as a constrained optimization problem in which the agent could reach the goal region while avoiding obstacles. They then reformulated this optimization problem using the epigraph form, a mathematical representation that could be solved using a deep reinforcement learning algorithm. This method overcame the limitations of existing reinforcement learning approaches and enabled the researchers to derive mathematical expressions for the system and combine them with existing engineering techniques.
In testing their approach, the researchers conducted control experiments with different initial conditions. The method achieved stability for all trajectories while maintaining safety, outperforming previous approaches. In a scenario inspired by a Top Gun movie, the researchers simulated a jet aircraft flying through a narrow corridor near the ground. Their controller effectively stabilized the jet, preventing crashes or stalls and outperforming other baselines.
The researchers believe this breakthrough technique has promising applications in designing controllers for highly dynamic robots that require safety and stability guarantees, such as autonomous delivery drones. It could also be used to assist drivers in navigating hazardous conditions by reestablishing stability when a car skids on a snowy road.
The researchers envision providing reinforcement learning with safety and stability guarantees for mission-critical systems. This approach represents a significant step toward achieving that goal. The team plans to improve the technique by accounting for uncertainty when solving the optimization and assessing its performance in real-world scenarios.
Experts have commended the MIT team for improving reinforcement learning performance in systems where safety is paramount. This breakthrough has far-reaching implications for generating safe controllers for complex scenarios, including a nonlinear jet aircraft model.