Neural networks have revolutionized how engineers design controllers for robots, leading to more adaptive and efficient machines. However, the complexity that gives these machine-learning systems their power also makes it challenging to ensure that a robot will safely and reliably complete its tasks.
Traditionally, safety and stability in control systems are verified using Lyapunov functions. If a Lyapunov function can be found such that its value consistently decreases, it indicates that the system will avoid unsafe or unstable conditions. However, applying these verification techniques to robots controlled by neural networks has been problematic due to the complexity of these systems.
Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and other institutions have developed new methods to rigorously certify Lyapunov functions in complex systems. Their algorithm efficiently searches for and verifies a Lyapunov function, providing a stability guarantee for the system. This approach promises safer deployment of robots and autonomous vehicles, including aircraft and spacecraft.
The researchers improved upon previous algorithms by using a cost-effective shortcut in the training and verification process. They generated counterexamples, such as adversarial sensor data that could disrupt the controller, and then optimized the robotic system to handle these scenarios. This approach helped machines learn to manage challenging conditions, ensuring safer operation across a broader range of environments. Additionally, the team developed a novel verification formulation that utilizes a scalable neural network verifier, α,β-CROWN, to provide rigorous worst-case scenario guarantees beyond the counterexamples.
“We’ve seen some impressive empirical performances in AI-controlled machines like humanoids and robotic dogs, but these AI controllers lack the formal guarantees that are crucial for safety-critical systems,” says Lujie Yang, an MIT electrical engineering and computer science (EECS) PhD student and CSAIL affiliate who co-authored a new paper on the project with Toyota Research Institute researcher Hongkai Dai SM ’12, PhD ’16. “Our work bridges the gap between that level of performance from neural network controllers and the safety guarantees needed to deploy more complex neural network controllers in the real world,” notes Yang.
In a digital demonstration, the team simulated a quadrotor drone equipped with lidar sensors stabilizing in a two-dimensional environment. Their algorithm successfully guided the drone to a stable hover position using the limited environmental information from the lidar sensors. In two other experiments, their approach enabled the stable operation of two simulated robotic systems under a wider range of conditions: an inverted pendulum and a path-tracking vehicle. These experiments, while modest, represent a significant advancement for the neural network verification community, especially since they included sensor models.
