Researchers Try to Give Robots Social Skills to Improve Interaction
Robots in movies are known for helping (or getting in the way of) people, but in the real world their social skills are quite limited. According to a new paper, computer scientists at MIT CSAIL are trying to fix this by teaching robots to interact with other robots and with humans in pursuit of their own goals.
“Robots will soon be everywhere in our world, and they really need to learn to communicate with us on human terms. They need to understand when to help and when to watch and see what they could do to keep something from happening”
The team calls this research the "first very serious attempt to understand what human-machine social interaction means." The problem they are trying to solve at a very early stage will undoubtedly receive more and more attention from roboticists as robots begin to play a huge role in our lives.
"Although machines and robots must interact with people, endowing them with social skills has been largely ignored. This is mainly due to the fact that tasks such as navigation, following commands and even playing are well defined, while social reasoning remains largely a theoretical problem. We demonstrate how social interactions can be effectively incorporated into MDPs (Markov decision processes) by recursively reasoning about the goals of other agents," the study says.
The research could lead to better human-robot interaction, for example in service homes, and even help psychologists better measure social interactions between people. To study these interactions, the researchers created a simulated 2D environment that allowed virtual robots to pursue both social and physical goals.
“We have discovered a new mathematical framework for modeling social interaction between two agents. If you are a robot and want to reach point X, and I am another robot and I see that you are trying to get to point X, I can cooperate by helping you get to point X faster. This could mean moving X closer to you, finding another better X, or performing any actions you would have taken at X. Our formulation lets the plan discover the "how," and we define the "what" in terms of the mathematical meaning of social interactions”
For example, a physical goal might be to go to a tree at a certain grid point, while a social goal is to guess what another robot is trying to do and then act on that, such as helping the other robot water the tree. A robot is rewarded for actions that bring it closer to its goals, with a corresponding reward for helping and an opposite reward for obstructing.
The team created three types of robots: the first pursues only physical goals, the second pursues physical and social goals but assumes that all robots pursue only physical goals, and the third assumes that everyone else has both social and physical goals, so it can take more complex actions, such as joining forces with others to achieve a goal.
Even very young babies seem to understand social interactions such as helping and obstructing, but we don't yet have machines that can perform this kind of reasoning with anything like human-level flexibility. The team created 98 different scenarios with all three types of robots. Twelve people watched about 200 video clips of robot interactions and then had to rate the physical and social goals.
The researchers reported: "In most cases, their model matched what people thought about the social interactions taking place in each frame." The researchers hope the results will serve as a "benchmark" that allows others to work on similar social interactions. The ultimate goal is not just to teach robots how to better interact in society, but to "delve into the human side of this."
“Can we run an objective test of your ability to recognize social interactions? Maybe there is a way to teach people to recognize these social interactions and improve their abilities”
Currently, the model is a relatively simple two-dimensional simulation. The research team is working on moving to a 3D version by adding a neural network-based robot planner to increase the speed at which robots learn from these actions.
Having an accurate mathematical model for addressing questions about social interactions has both practical value, as one can imagine chatbots and robots driven by social MDPs, and theoretical value, if MDP tools that have so successfully organized navigation research are used to shed light on which social interactions are truly of extreme importance for human well-being and human civilization.
