- MIT researchers have developed a new AI-based method for controlling autonomous robots, addressing the “stabilize-avoid” challenge.
- The approach could potentially be applied in situations where dynamic robots must meet safety and stability requirements, like autonomous drones.
A novel AI technique to govern autonomous robots has been developed by researchers at the Massachusetts Institute of Technology (MIT). The new method offers a solution to the complex “stabilize-avoid” problem, a long-standing conflict that most existing AI methods fail to reconcile. Through the MIT’s approach, not only is there a marked increase in stability, but it also maintains or enhances the safety compared to other prevalent techniques.
Typically, the majority of techniques aimed at solving stabilize-avoid problems attempt to simplify the system, rendering it solvable through rudimentary mathematical methods. However, these simplified solutions often falter when exposed to real-world dynamics.
The video shows the successful application of the method: Piloting a simulated jet in a complex low-altitude scenario. Credits: Massachusetts Institute of Technology.
Credits: Massachusetts Institute of Technology.
The team from MIT deconstructed the problem into two steps. The first reframed the stabilize-avoid problem as a constrained optimization problem, allowing the autonomous agent to reach and stabilize its goal while ensuring obstacle avoidance through applied constraints. The second step transformed the constrained optimization problem into an epigraph form and solved it using a deep reinforcement learning algorithm, a technique that allowed them to circumvent the difficulties other methods encountered with reinforcement learning.
Practical Application and Testing
The new technique was tested under a variety of control experiments, each with distinct initial conditions. For instance, in some simulations, the autonomous agent had to reach and remain within a goal region while executing sharp maneuvers to bypass obstacles in its path. When tested against various baselines, the MIT approach successfully stabilized all trajectories while maintaining safety.
The research indicates that this method could be utilized as a foundation for developing controllers for dynamic robots that require strict safety and stability standards, like autonomous delivery drones. Furthermore, it could be part of larger systems where it is activated under certain conditions to ensure safety, such as in an automobile skidding on a snowy road.