It's important to make sure robots can interact with humans in the best way possible. They need to not only be efficient and safe, but also be intuitive and explainable. One crucial aspect is allowing people to customize and influence a robot's behavior. For example, when robots move around near humans, they should adapt to social rules, context, and most importantly people's preferences. So, how can we easily teach a robot what we want and while making sure it follows important rules like reaching its goal and avoiding collisions?

We actively work on learning architectures and feedback interfaces that enable preference-aligned robot behavior. For example, we have developed an immersive virtual reality interface, where the user intuitively demonstrates preferences. From those demonstrations, the robot can learn a personalized navigation controller within a safe computer simulation, kind of like how we learn from trial and error. We could show that the robot controlled by our system made people feel more comfortable when interacting with it.