Engineers at MIT have developed an approach that can be used to identify potential failures in autonomous systems before they are deployed. The approach, which can be paired with any autonomous system, uses an automated sampling algorithm to quickly identify a range of likely failures and suggest repairs to avoid system breakdowns.
The algorithm takes a different approach from other automated searches, which typically focus on spotting the most severe failures. The MIT team found that these approaches could miss subtler, yet significant vulnerabilities that their algorithm was able to catch. The new algorithm is designed to identify a diversity of failures rather than condensing them into the most severe or likely failure.
The inspiration for this research came from a major power grid failure that occurred in Texas in 2021. The unexpected failure of the power grid during winter storms left millions of homes and businesses without power for days. This led the MIT engineers to question whether such failures could be predicted beforehand and whether steps could be taken to strengthen system vulnerabilities.
To test their approach, the team simulated various autonomous systems, including power grids, aircraft collision avoidance systems, rescue drones, and a robotic manipulator. In each case, the algorithm quickly identified potential failures and suggested repairs to prevent them.
The algorithm works by generating random changes within a system and assessing the system’s sensitivity to those changes. The more sensitive a system is to a particular change, the more likely that change is associated with a possible failure. This approach allows the algorithm to uncover a wider range of failures and identify fixes by tracing back the chain of changes that led to a specific failure.
The researchers demonstrated the effectiveness of their approach on a robotic manipulator by running simulations and then testing the algorithm’s predictions in a real-world scenario. The algorithm accurately predicted how the robot would fail and suggested a fix that allowed the robot to successfully complete its task.
The team believes that their approach has the potential to improve the resilience of autonomous systems in various industries. By identifying and addressing potential failures before deployment, these systems can be made more reliable and trustworthy.
The researchers presented their findings at the Conference on Robotic Learning, highlighting the importance of understanding the limits of autonomous systems and knowing when and how they are likely to fail. By taking a proactive approach to failure detection and prevention, engineers can mitigate the risks associated with autonomous systems and ensure their dependable operation.
1. Source: Coherent Market Insights, Public sources, Desk research
2. We have leveraged AI tools to mine information and compile it