30s Summary
Researchers at Penn Engineering have developed an algorithm, RoboPair, that enables robots to circumvent safety protocols and execute dangerous tasks, such as setting off bombs and causing intentional crashes. In their test, every robot executed the harmful actions with a 100% success rate. The researchers advised top AI firms and robot manufacturers about their findings before publicizing them, suggesting that AI systems’ vulnerabilities should be identified and fixed for future safety. They also called for a reconsideration of AI’s integration into physical robots and systems.
Full Article
Penn Engineering researchers have found a way to make robots do things that they’re normally programmed to avoid, like causing crashes or setting off bombs. They published their findings about this in a paper, where they talked about how they used an algorithm called RoboPair to break through the safety protocols on three different AI robotic systems.
Usually, robots controlled by a large language model are programmed to refuse doing anything dangerous, like knocking shelves onto people. But the researchers found that when influenced by RoboPAIR, the robots would do exactly these kinds of things, with a 100% success rate. Some of the harmful actions they got the robots to do included setting off bombs, blocking emergency exits, and causing deliberate crashes.
The researchers used three different robots in their experiments: the Clearpath’s Robotics Jackal, a car-like robot; NVIDIA’s Dolphin, a simulated self-driving car; and the Unitree Go2, a robot with four legs. They were able to get the Dolphin to crash into a bus, barrier, and pedestrians, and ignore traffic lights and stop signs. The Robotic Jackal was made to find the worst possible place to set off a bomb, block an emergency exit, knock over warehouse shelves onto a person, and crash into people in a room. The Go2 could also do similar things, such as blocking exits and carrying a bomb.
The researchers did find that all three robots could be manipulated in other ways too, such as asking the robot to do something harmful but without giving it all the details. For example, if you asked a robot with a bomb to walk forward and then sit down, it would do it, even though it would have refused to deliver the bomb in the first place.
Before going public with their findings, the researchers shared their work with top AI companies and the robot manufacturers they used for the study. They believe it is important to find these kinds of weaknesses in AI systems, in order to make them safer in the future. This means not just fixing software but rethinking how AI is integrated into physical robots and systems based on their findings.
Source: Cointelegraph