Computer Vision News - June 2021

of these technologies really function independently. They are a part of a complex system. We co-design them with our social norms and policy as well, so that we can achieve a large- scale introduction of these systems. There’s a lot we can learn from aviation. For example, things that you think should be clumsy aspects of interaction with a system, maybe the interface doesn’t allow you to navigate to find the right window quickly, to understand some particular subsystem. Building a person’s mental model of a system is critical. We’ve had many aviation accidents. There was an interplay of this effect with the Boeing 737 MAX mode confusion and not understanding the current state of the system, how the actions will impact the behavior of that system. The mode confusion has resulted inmany aviation accidents. We actually see it in the interaction with drivers, with intelligent vehicles like the Tesla. These are challenges that are not just specific to aviation, but the lessons that we’ve learned there can translate and can help us design systems that are safer for us broadly in society. In general, what are we doing right and what are we doing wrong as humans interacting with machines? I spent much of my current career in robotics for manufacturing, helping to deploy collaborative robots that work alongside people who build planes and build cars. The interesting thing about deploying a robotic system that moves and works around people is that largely with these systems, they just see people as obstacles. We see this with sidewalk delivery robots or robots in grocery stores. We’re really no different than any other obstacle that systems might encounter, and that’s a major limitation for these systems to integrate effectively with people. Just like the robot needs to understand what other cars are, what trees are, they need to understand more about people than us just being obstacles in their environment. So that’s a key focus of my lab’s research. Can you tell us a bit more about robots in healthcare, like in robotic surgery? This is a very exciting time in the introduction of technology and AI in healthcare. We have really excellent examples of research at MIT that can help diagnose cancer in images. What we see there and what we know from many other contexts is that humans and machines in AI have different capabilities. The errors that humans make are going to be different from the errors that an AI system will make. By cleverly architecting the system, a system that enhances human capability rather than replacing it, that’s where you can achieve something better 28 Women in Science