Artificial Social Intelligence Challenges, Environments, and Mechanisms

  Artificial intelligence (AI) applications are becoming increasingly com mon in humans’ daily lives. However, to be truly useful for us, AI need to  be not only task solvers, but also decision makers that interact with us and  other AI agents intelligently in the grounded world.  

This thesis studies this capability of AI, defining the concept of artificial  social intelligence. Humans’ social intelligence emphasizes relationship man agement in an interpersonal context, while artificial social intelligence covers  a wider range of capabilities, from rudimentary ones, e.g. maintaining memory and self-consistency, to more sophisticated ones, e.g. realizing long-term  goals in social and embodied interactions.  

There are three key challenges in achieving social intelligence:  

1. Safe and strategic decision making in social interactions: to identify and  tackle these challenges, this thesis will introduce a platform that makes  it easy to develop and deploy socially intelligent AI agents in automated  environments, called Sotopia. Sotopia provides the mechanisms for  training and evaluating agents through social interactions.  
2. Understanding humans’ intention and mental models: this thesis will  also present the computational implementation of the core mechanism  of socially intelligent agents: Theory of Mind (ToM). I will show how  ToM is useful for quickly adapting to new interlocutors, and even helps  language agents acquire language skills through social interactions. 
3. Grounding social interactions in the physical world: finally, this thesis  will discuss two ways to ground social interactions. The first is to  ground them in the virtual and embodied world, and the second is  to ground them in the physical world. Following the first way, this  thesis introduces two benchmarks that test AI agents’ abilities to follow  instructions and answer grounded questions. And following the second  way, I will demonstrate Sotopia-Robots, an extension of Sotopia that  highlights realistic human-robot interaction.  

This thesis studies a new field for AI research– social intelligence which intersects with various traditional AI fields: Cognitive Science, Natural Language Processing (NLP), and Robotics. Through identifying the  various challenges, I build up the prototype of an ecosystem that can unify  vastly different environments and allow for future research advances on the  training, evaluation, and deployment of socially intelligent AI in the real  world.