Photo credit: news.mit.edu

New Techniques Enhance Safety and Performance in AI-Controlled Robots

Neural networks have revolutionized the engineering of robotic controllers, yielding machines that are more adaptive and efficient. However, the complexity that makes these brain-like systems powerful also introduces challenges in ensuring that robots using them perform tasks safely.

Engineers traditionally rely on Lyapunov functions to verify the safety and stability of controlled systems. If a Lyapunov function’s value decreases consistently, unsafe or unstable states are avoided. However, existing methods for verifying Lyapunov conditions don’t scale well to the complex machines controlled by neural networks.

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and other institutions have developed new techniques to rigorously certify Lyapunov calculations in sophisticated systems. Their algorithm efficiently identifies and verifies a Lyapunov function, providing stability guarantees. This advancement promises safer deployments of robots and autonomous vehicles, including aircraft and spacecraft.

To improve upon previous algorithms, the researchers employed cost-effective counterexamples, such as adversarial sensor data that could disrupt the controller, then optimized the system to manage these scenarios. This strategy helped machines handle challenging situations, extending their safe operational range. Additionally, the team formulated a scalable verification method using the α,β-CROWN neural network verifier to provide rigorous guarantees beyond counterexamples.

“AI-controlled machines like humanoids and robotic dogs have shown impressive empirical performance, but lack formal safety guarantees crucial for critical systems,” says Lujie Yang, MIT electrical engineering and computer science PhD student and CSAIL affiliate. “Our work bridges that performance gap by offering safety guarantees necessary for deploying complex neural network controllers in real-world applications,” adds Yang, co-lead author of the study with Toyota Research Institute researcher Hongkai Dai.

The team demonstrated their algorithm in simulated scenarios, guiding a quadrotor drone with lidar sensors to a stable hover position using only the limited data from these sensors. Their approach also stabilized two other simulated robotic systems over diverse conditions: an inverted pendulum and a path-tracking vehicle, marking a significant step forward for the neural network verification community.

“The use of neural networks as Lyapunov functions presents tough global optimization challenges that require scalability,” notes Sicun Gao, associate professor at the University of California, San Diego. “This work makes significant contributions by enhancing algorithmic approaches for neural Lyapunov functions in control problems, achieving notable improvements in scalability and solution quality,” says Gao, who was not involved in this research.

The stability approach developed by Yang and her colleagues has wide-ranging safety-critical applications, from autonomous vehicles to drones engaged in delivery or terrain mapping. The general nature of these techniques suggests potential utility in fields like biomedicine and industrial processing.

While their current method is scalable, the researchers aim to enhance performance in higher-dimensional systems and incorporate additional data types, such as images and point clouds.

Future research will focus on providing stability guarantees for systems in uncertain environments or facing disturbances, like a drone encountering a strong gust of wind. The team also seeks to apply their methods to optimization problems, aiming to minimize task completion time and distance while maintaining stability. They plan to extend their technique to real-world machines, such as humanoids, where stability during physical interaction is critical.

Russ Tedrake, Toyota Professor of EECS, Aeronautics and Astronautics, and Mechanical Engineering at MIT, and CSAIL member is a senior author of this research. The paper, which also credits Zhouxing Shi, Cho-Jui Hsieh, and Huan Zhang, is supported by Amazon, the National Science Foundation, the Office of Naval Research, and Schmidt Sciences’ AI2050 program. The researchers will present their findings at the 2024 International Conference on Machine Learning.

Source
news.mit.edu