Neural networks have had a huge impact on how engineers design robot controllers, spurring the development of more adaptive and efficient machines. However, brain-like machine learning systems are a double-edged sword: Their complexity makes them powerful, but it also makes it difficult to ensure that a neural-networked robot will complete its task safely.
The traditional way to check safety and stability is through techniques called Lyapunov functions. If you can find a Lyapunov function that is constantly decreasing in value, you can tell that unsafe or unstable situations associated with higher values will never occur. However, for robots controlled by neural networks, previous methods for checking Lyapunov conditions have not been suitable for complex machines.
Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory and elsewhere have developed new techniques that rigorously verify the correctness of Lyapunov calculations in more complex systems. Their algorithm efficiently searches for and verifies the Lyapunov function, providing a guarantee of the system’s stability. The approach could potentially enable safer deployment of robots and autonomous vehicles, including aircraft and spacecraft.
To outperform previous algorithms, the researchers found an economic shortcut to the training and verification process. They created cheaper counterexamples—for example, hostile data from sensors that could have thrown off the controller—and then optimized the automated system to account for these examples. Understanding these edge cases helped the machines learn how to handle difficult conditions, enabling them to operate safely in a wider range of circumstances than was previously possible. They then developed a new verification formula that enables a scalable neural network verifier, α,β-CROWN, to provide stringent guarantees for worst-case scenarios beyond the counterexamples.
“We’ve seen some impressive experimental performance in AI-controlled machines like humans and robotic dogs, but these AI controllers lack the formal guarantees that are critical for safety-critical systems,” says Lujie Yang, an MIT EECS doctoral student and CSAIL member 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 this level of performance from neural network controllers and the safety guarantees needed to deploy more complex neural network controllers in the real world,” Yang notes.
In a digital demonstration, the team simulated how a quadcopter drone equipped with lidar sensors stabilizes itself in a two-dimensional environment. Their algorithm successfully guided the drone into a stable hovering position, using only the limited environmental information provided by the lidar sensors. In two other experiments, their approach enabled stable operation of two simulated robotic systems over a wider range of conditions: an inverted pendulum and a path-following vehicle. These experiments, while modest, are relatively more complex than what the neural network verification community has been able to do before, especially since they involved sensor models.
“Unlike common machine learning problems, the accurate use of neural networks as Lyapunov functions requires solving difficult global optimization problems, and thus scalability is a major bottleneck,” says Sikon Gao, an associate professor of computer science and engineering at UC San Diego, who was not involved in the work. “The current work makes an important contribution by developing algorithmic approaches that are better designed for the specific use of neural networks as Lyapunov functions in control problems. It achieves a dramatic improvement in scalability and solution quality compared to existing approaches. The work opens up exciting directions for further development of optimization algorithms for neural Lyapunov methods and the accurate use of deep learning in control and robotics in general.”
The stabilization approach Yang and her colleagues have potential applications where safety is of the utmost importance. It could help ensure smoother flight for autonomous vehicles, such as airplanes and spacecraft. Likewise, drones that deliver items or map different terrains could benefit from such safety assurances.
The technologies developed here are very general and are not limited to robotics; the same technologies could help in other applications, such as biomedicine and industrial processing, in the future.
While the technique represents an improvement over previous work in terms of scalability, the researchers are exploring how it could work better in higher-dimensional systems. They also want to consider data beyond lidar readings, such as images and point clouds.
As a future research direction, the team wants to provide the same stability guarantees for systems in uncertain, turbulent environments. For example, if a drone encounters a strong gust of wind, Yang and her colleagues want to ensure that it can stay in the air and complete the required mission.
They also plan to apply their method to optimization problems, where the goal is to minimize the time and distance a robot needs to complete a task while remaining stable. They plan to extend their technique to humans and other machines in the real world, where a robot needs to remain stable while communicating with its surroundings.
Ross Tedrick, MIT professor of electronics, mechanical, and aerospace engineering, vice president for robotics research at the TRACE Institute, and a member of the Institute for Computer Science and Artificial Intelligence, is a senior author of the paper. The paper also credits UCLA doctoral student Chuxing Shi and assistant professor Chu-Jui Hsieh, as well as assistant professor Huan Zhang of the University of Illinois at Urbana-Champaign. Their work was supported in part by Amazon, the National Science Foundation, the Office of Naval Research, and the AI2050 program at Schmidt Sciences. The researchers’ paper will be presented at the 2024 International Conference on Machine Learning.
